# Auto-GPT

> Source: https://aiwiki.ai/wiki/autogpt
> Updated: 2026-06-23
> Categories: AI Agents, Open Source AI
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

**AutoGPT** (also stylized **Auto-GPT**) is an open-source autonomous [AI agent](/wiki/ai_agents) framework that uses [large language models](/wiki/large_language_model) (LLMs), primarily [OpenAI](/wiki/openai)'s [GPT-4](/wiki/gpt-4), to accomplish user-defined goals with minimal human intervention. It was released on March 30, 2023, by British game developer Toran Bruce Richards, two weeks after OpenAI announced [GPT-4](/wiki/gpt-4), and within thirteen days had collected 30,000 GitHub stars, crossing 100,000 stars a few weeks later in what was widely described as the fastest-growing open-source project in GitHub history at the time.[^1][^2][^3] The project is maintained by Significant Gravitas Ltd., a UK-based company that raised $12 million in venture funding in October 2023 from Redpoint Ventures and GitHub.[^4][^5]

Unlike a [chatbot](/wiki/chatgpt) that responds to one prompt at a time, AutoGPT operates through a self-prompting loop: the user supplies a high-level goal in natural language, and the agent breaks it into subtasks, picks tools (web browsing, file operations, code execution, API calls), evaluates each outcome, and iterates until it judges the objective complete or hits a step limit.[^6][^7][^8] The approach reframed how people interacted with LLMs in 2023, moving from prompt-response chat to goal-directed autonomous behavior, and it became the highest-profile early demonstration of what later writers would call [agentic AI](/wiki/agentic_ai).[^9][^10]

As of May 2026, AutoGPT has changed shape considerably from its original command-line script. The repository now ships an "AutoGPT Platform" with a low-code visual builder, a marketplace of pre-built agents, and integrations with dozens of external services, alongside the legacy "AutoGPT Classic" agent.[^11][^12] The GitHub repository has accumulated roughly 184,000 stars and 46,200 forks across more than 8,100 commits.[^1] Multiple 2026 framework comparisons place AutoGPT in the "experimental" or "specialized" tier rather than the production-ready tier, noting that systems such as [LangGraph](/wiki/langgraph), [CrewAI](/wiki/crewai), and Microsoft's [AutoGen](/wiki/autogen) have "largely eclipsed" it in active enterprise deployment.[^13][^14]

## When was AutoGPT released and who created it?

### Origins and creator

AutoGPT was created by Toran Bruce Richards, a 2020 graduate of Edinburgh Napier University with a Bachelor of Science in Game Development.[^15] Richards is the founder and CEO of Significant Gravitas Ltd., a small Edinburgh-based studio originally focused on applying software techniques from the video game industry to non-gaming use cases.[^15][^16] His earlier work included Orbits: The Sandbox, an N-body space simulation aimed at building physics intuition, that began as a 48-hour prototype at the Edinburgh Global Game Jam in 2020.[^15]

According to Richards' account on OpenUK and his interviews with Motherboard and Fortune, the project that became AutoGPT began as a personal experiment called "Entrepreneur-GPT" (sometimes written EnterpreneurGPT), in which he tried to give [GPT-4](/wiki/gpt-4) the ability to plan and execute multi-step business tasks without continuous user prompting.[^2][^15] Development was largely unnoticed on Twitter and GitHub until late March 2023, when Richards published a public demo as Auto-GPT.[^2][^15]

OpenAI had released GPT-4 on March 14, 2023; AutoGPT was committed publicly on March 30, 2023, making it one of the first widely visible attempts to wrap the newer model in an autonomous loop.[^1][^17] Richards has said the decision to release the code under a permissive license stemmed from a desire to democratize access to capable [LLM](/wiki/large_language_model) tooling and to prevent the autonomous-agent pattern from becoming the exclusive property of large labs.[^15]

### Viral growth (April 2023)

AutoGPT's release coincided with intense public interest in generative AI following the late-2022 launch of [ChatGPT](/wiki/chatgpt). On April 3, 2023, Auto-GPT became the top trending repository on GitHub.[^2] By April 12, 2023, it had reached 30,000 stars; within a few more weeks it crossed 100,000 stars, briefly overtaking projects such as PyTorch on the star leaderboard despite being only weeks old.[^2][^9] Multiple commentators called it the fastest-growing open-source project on GitHub up to that point.[^9][^10]

On April 2, 2023, [Andrej Karpathy](/wiki/andrej_karpathy), then formally a co-founder of [OpenAI](/wiki/openai), wrote on Twitter that "AutoGPTs" were the "next frontier of prompt engineering" and described 1 GPT call as "just like 1 instruction on a computer" that could be "strung together into programs."[^17][^18] The post amassed more than 4,700 likes within ten days and is widely credited with pulling additional attention to the project from machine-learning researchers and venture investors.[^17]

Mainstream technology coverage followed quickly. Motherboard (Vice) published an article on April 4, 2023, in which Richards described AutoGPT as turning a large language model "from what is essentially an advanced auto-complete into an independent agent capable of carrying out actions and learning from its mistakes."[^2][^19] Fortune ran a feature on April 15, 2023, framing AutoGPT and BabyAGI as a new category of "GPT-based agents" buzzing through Silicon Valley.[^20] TechCrunch followed on April 22, 2023, describing the system as GPT-3.5 and GPT-4 "paired with a companion bot that instructs GPT-3.5 and GPT-4 what to do."[^21] Will Knight covered the project critically for Wired, reporting that AutoGPT failed at relatively simple research tasks, including locating a public figure's email address.[^2][^22] The AutoGPT hashtag also trended worldwide on Twitter, with viral demonstrations including agents that drafted podcast outlines, built simple web applications, and conducted competitor research without manual intervention.[^9][^23]

### Funding

On October 17, 2023, Significant Gravitas Ltd. announced a $12 million venture round led by Redpoint Ventures, with participation from GitHub through GitHub Fund.[^4][^5] The Hacker News thread on the funding cited the company's plan to use the capital to build a hosted product and a marketplace, expand the engineering team, and accelerate development of the agent framework.[^24] The company remained registered in the United Kingdom despite the U.S. investor base.[^5][^16]

### Evolution to a platform (2023 to 2026)

The project's roadmap can be divided into three eras: the original Python script (2023), a refactored classic agent with the Agent Protocol (late 2023 through 2024), and the AutoGPT Platform with visual workflow blocks (2024 onward).[^9][^11]

| Version | Date | Key changes |
| --- | --- | --- |
| v0.1.0 to v0.4.x | March to November 2023 | Original command-line agent with self-prompting loop, web browsing, file operations, plugin system; first-party plugins added April 2023[^1][^25] |
| v0.5.0 | December 14, 2023 | Major refactoring of the classic agent. Added Agent Protocol REST API, new Agent UI, expanded storage options (S3, Google Cloud Storage), and improved CLI/TTY mode with state save and resume[^26] |
| v0.5.1 | April 2024 | Incremental improvements to the classic [AutoGPT agent](/wiki/autogpt_agent) before the platform pivot[^1] |
| Platform rewrite | July 22, 2024 | Visual workflow builder and modular backend introduced; agents restructured as composable "blocks"; vector-database backends replaced with a JSON file plus NumPy similarity search[^9][^27] |
| Platform launch | February 5, 2026 | Public introduction of the AutoGPT Platform with AutoGPT Server, AutoGPT Builder, and a marketplace of pre-built agents[^11] |
| Platform Beta v0.6.57 | April 22, 2026 | Added Agent Briefing Panel, subscription-tier billing via Stripe Checkout, xAI Grok 4.20 models from OpenRouter, follow-up message queueing, extended thinking streaming[^28] |
| Platform Beta v0.6.58 | April 29, 2026 | Added [Claude](/wiki/claude) Opus 4.7 model support, Settings v2 redesign, Redis Cluster client support, Web Push notifications via VAPID, dynamic BlockCostType billing system[^28] |
| Platform Beta v0.6.59 | May 7, 2026 | Settings v2 billing page with subscription and automation credits, tier-based workspace file storage limits, admin CSV exports, yearly billing[^28] |
| Platform Beta v0.6.60 | May 13, 2026 | "Trigger On Anything" workflow flexibility, Slack and Discord integrations, smart bot-to-bot communication, AutoPilot task queue, Stripe automatic tax[^28] |
| Platform Beta v0.6.61 | May 20, 2026 | Chat search modal, session sidebar pagination, profile popover redesign, "Response stopped" banner, WCAG AA accessibility for inputs[^28] |

The transition from "AutoGPT Classic," a single autonomous Python agent, to the "AutoGPT Platform," a web-based system with visual block composition, marked the project's most significant architectural shift. The classic agent remains in the repository under the MIT License but is no longer the primary direction of development.[^11][^12]

### The post-hype era

The autonomous-agent boom that AutoGPT helped trigger cooled substantially over 2024. By mid-2024 multiple commentators observed that "fully autonomous" general-purpose agents had not delivered on their early 2023 promises, with the dominant explanations being looping behavior, hallucinated steps, and the compounding of small per-step error rates over long task sequences.[^9][^29] Yohei Nakajima archived BabyAGI in September 2024, snapshotting the original repository and relaunching the project as an explicit research sandbox.[^30] AgentGPT's parent company Reworkd raised $4 million in 2024 and pivoted to web-scraping infrastructure, archiving the original AgentGPT repository.[^31] AutoGPT itself underwent its July 2024 visual-builder rewrite, which one retrospective described as "a completely different product from the one that earned those 100,000 stars."[^9]

By 2025-2026, the production-oriented end of the agent market had shifted toward systems that were explicitly less autonomous: IDE-anchored coding assistants such as [Cursor](/wiki/cursor) and [Claude Code](/wiki/claude_code), task-delegation systems such as [Devin](/wiki/devin), and multi-agent frameworks such as [LangGraph](/wiki/langgraph), [CrewAI](/wiki/crewai), and [AutoGen](/wiki/autogen) that emphasized structured roles, defined handoffs, and human-in-the-loop checkpoints rather than open-ended self-prompting.[^13][^14][^32]

## How does AutoGPT work?

### The autonomous agent loop

The core of the classic AutoGPT is an iterative loop, sometimes summarized as plan, act, reflect, around a single underlying [LLM](/wiki/large_language_model).[^6][^7][^33] In its standard configuration the loop proceeds as follows:

1. **Goal definition.** The user provides a high-level objective in natural language (for example, "Research the top five competitors in the electric vehicle market and produce a one-page summary").
2. **Task decomposition.** The agent prompts the LLM to break the goal into smaller subtasks and order them, building an internal task list.
3. **Self-prompting.** Rather than waiting for the user, the agent constructs its own prompts based on the current task, recent actions, and observed outcomes, drawing on a contract-style system prompt.[^34]
4. **Action execution.** The agent picks one of a small set of available tools (web search, page fetch, file read or write, Python execution, API call) and runs it.
5. **Evaluation and criticism.** After each action, the agent asks the LLM to review the result against the active subtask and the overall goal, decide whether to continue, abandon, or revise the current approach, and produce structured JSON output.[^8][^34]
6. **Iteration.** Steps 3 through 5 repeat until the agent judges the goal complete or hits a configured step or budget limit.

The autonomous loop in AutoGPT is an instance of the broader ReAct pattern (reason and act) introduced by Yao et al. in 2022, in which a model interleaves chain-of-thought reasoning with tool actions and observations.[^33][^35] Later research, notably the [Reflexion](/wiki/reflexion) paper from Noah Shinn and collaborators in March 2023, formalized the addition of post-hoc "verbal" self-reflection on past trials, which became a recurring pattern in agent design.[^36][^37] AutoGPT predates Reflexion's release by ten days and does not implement the full Reflexion algorithm; it does include an explicit criticism step that broadly resembles short-horizon reflection.[^33][^36]

### System prompt and configuration

The classic AutoGPT system prompt is structured as a contract that lists the agent's name, role, constraints, available commands and resources, evaluation criteria, and preferred best practices.[^34][^38] Each agent is configured with five user-supplied parameters:[^6][^38]

1. A name for identifying the agent
2. A role description of the agent's purpose
3. A list of constraints
4. A list of available resources (tools and information sources)
5. A list of best practices to follow

The model is also instructed to respond in a fixed JSON schema with keys for thoughts, reasoning, plan, criticism, and the next command. Format parsing of this JSON output represents a significant portion of the original AutoGPT code base; Lilian Weng's June 2023 survey "LLM-Powered Autonomous Agents" noted that "a lot of code in AutoGPT [is] devoted to format parsing."[^33]

### Memory architecture

AutoGPT's memory subsystem splits between short-term and long-term storage:[^6][^8]

- **Short-term memory** holds recent messages, session state, and the outputs of the most recent actions in an in-context buffer that is included on each LLM call.
- **Long-term memory** persists information across sessions. Early versions exposed several pluggable backends: local file storage, [Pinecone](/wiki/pinecone) (vector database), Redis, Milvus, and Weaviate, with embedding-based semantic search used to retrieve relevant past content into the prompt.[^39][^40]

The retrieval pattern resembled that used in [retrieval-augmented generation](/wiki/retrieval_augmented_generation), with [embeddings](/wiki/embeddings) representing prior actions and observations and similarity search returning a handful of relevant items per step. The intent was to mitigate the [context window](/wiki/context_window) limits of GPT-3.5 and GPT-4 by avoiding replaying the entire history on each call.[^8][^40]

In practice, AutoGPT engineers later removed support for external [vector databases](/wiki/vector_database) (Pinecone, Milvus, Redis, Weaviate). Internal experiments and external commentary found that agent runs typically did not generate enough distinct factual content to justify a managed vector store, and the team replaced the architecture with a simple JSON file and NumPy similarity operations.[^9][^39]

### Tool integration

AutoGPT agents interact with the world through a configurable command set:[^6][^7]

| Tool category | Capabilities |
| --- | --- |
| Web browsing | Search engine queries, page fetching, scraping, data extraction |
| File operations | Read, write, modify local files; manage workspaces |
| Code execution | Author and run Python in a sandboxed environment |
| API integration | Issue HTTP requests, send emails, post to messaging platforms |
| Image generation | Call [text-to-image](/wiki/text_to_image) models (added in platform versions) |

These tools allow a single agent to research, analyze, write, and act inside one autonomous workflow.

### Architecture components

The classic AutoGPT is a Python application that wraps the [OpenAI API](/wiki/openai_api). It consists of:[^41][^42]

- **The agent core**, which manages the autonomous loop, prompt construction, decision making, and JSON parsing
- **A plugin system** that extends functionality through first-party and third-party plugins
- **A CLI** for configuration, agent launch, and progress monitoring
- **An optional web-based GUI** for users who prefer a graphical interface

It requires Python 3.10 or later, an OpenAI API key, and either a direct local install or a Docker setup.[^42] In its 2023 incarnation, AutoGPT defaulted to GPT-3.5 with optional GPT-4 calls for the planning step, primarily for cost reasons; users could override the configuration to use GPT-4 throughout.[^25][^43]

## What is the AutoGPT Platform?

### Architecture

The AutoGPT Platform, publicly described in agpt.co's "Introducing the AutoGPT Platform" announcement, is a two-part system:[^11][^12]

**AutoGPT Server** holds the core agent and orchestration logic, the marketplace, and the infrastructure for running agents on the server side. It exposes APIs for agent execution, scheduling, and integration with external services.[^11]

**AutoGPT Frontend** (the AutoGPT Builder) is a low-code visual interface for designing and configuring agents. Users drag blocks onto a canvas, connect their inputs and outputs, and configure their parameters. The frontend also provides deployment controls, dashboards, and analytics.[^11][^44]

### Blocks

Blocks are the fundamental composable units of the AutoGPT Platform. Each block represents an individual action or integration with inputs, outputs, and a transformation function. Block categories include:[^44]

- **External services** such as Slack, Discord, GitHub, Google services, Reddit, email systems, and CRMs[^28]
- **Data processing and transformation** for parsing, filtering, summarizing, and shaping data
- **LLM integrations** for [OpenAI](/wiki/openai), [Anthropic](/wiki/anthropic) (including [Claude](/wiki/claude) Opus 4.7 since v0.6.58), Groq, Llama, and xAI Grok models[^28]
- **Custom scripts** for user-supplied logic
- **Conditional logic** for branching and control flow

This shift from a single recursive loop to a graph of typed blocks represents a fundamental architectural change: rather than having the LLM decide every next step at run time, builders now wire together a graph of predictable units, with LLM calls embedded as a particular kind of block.[^9][^11]

### Marketplace and credits

The AutoGPT Platform marketplace allows users to discover, share, and deploy pre-built agents for tasks such as lead generation, SEO content writing, customer support automation, and meeting preparation.[^11] Execution is billed in platform credits, with tier-based subscription pricing through Stripe Checkout introduced in v0.6.57 and refined through v0.6.60.[^28]

### AutoGPT Forge

Forge is a toolkit, originally hosted as a standalone repository and later folded into the main AutoGPT codebase, that provides boilerplate for building custom agent applications. It handles common infrastructure (API endpoints, memory management, tool interfaces) so developers can focus on agent-specific logic. Forge implements the Agent Protocol, ensuring interoperability with other agent frameworks and benchmark harnesses.[^45][^46] The original Auto-GPT-Forge repository was archived as read-only on September 13, 2023, after its components were merged into the main project.[^46]

### AGBenchmark

AGBenchmark is AutoGPT's performance evaluation framework, designed to test any agent that implements the Agent Protocol. It includes:[^1][^45]

- A library of challenges defined using a JSON-based domain-specific language
- Categorized test suites that developers can customize
- Real-time progress tracking with pass, fail, or in-progress status
- Integration with the AutoGPT Arena Leaderboard

The benchmark is framework-agnostic by design: any agent that exposes the Agent Protocol HTTP interface can be evaluated under AGBenchmark.[^45]

### Agent Protocol

AutoGPT, working with the AI Engineer Foundation, helped define and promote the Agent Protocol, an open API specification for communicating with autonomous agents.[^45] The protocol is defined using an OpenAPI specification and provides a uniform REST interface for tasks such as submitting goals, retrieving outputs, and managing agent state, regardless of the underlying framework, language, or hosting model.[^45] Its goals included interoperability between agents from different projects, simplified cross-framework benchmarking, and easier integration of agents into downstream applications. The Agent Protocol remains distinct from the later [Agent2Agent (A2A) protocol](/wiki/a2a_protocol) introduced by Google in 2025 for multi-agent coordination.

### AutoGPT Arena hackathon

In 2023, the AutoGPT team partnered with lablab.ai on the AutoGPT Arena Hacks competition. The four-week event drew more than 5,000 participants in about 500 teams, who were tasked with building agents that performed on a comprehensive AI agent benchmark.[^47] Challenges fell into categories such as "Scrape and Synthesize" (web data extraction and dataset creation) and "Data Mastery" (data science tasks), with most agents performing well in only one category.[^47] The Arena Leaderboard recognized evo.ninja, an open-source generalist agent, as the top-scoring entry.[^48]

## System requirements

Running AutoGPT locally requires:[^42]

| Component | Minimum | Recommended |
| --- | --- | --- |
| CPU | 2 cores | 4+ cores |
| RAM | 8 GB | 16 GB |
| Storage | 10 GB free | 10 GB+ free |
| Docker Engine | 20.10.0+ | Latest stable |
| Docker Compose | 2.0.0+ | Latest stable |
| Node.js | 16.x+ | Latest LTS |
| Python | 3.10+ | 3.11+ |
| OpenAI API | Paid account | Paid account with billing configured |

Git is required for cloning the repository, and npm 8.x or later is required for the frontend components. A paid OpenAI account is recommended over free-tier accounts, since free-tier rate limits (three API calls per minute) tend to crash the autonomous loop.[^42]

## Deployment options

AutoGPT supports two primary deployment paths:[^1][^11]

- **Self-hosted.** Users download the source from GitHub and run the platform on their own infrastructure, retaining control of data and configuration but assuming setup and maintenance overhead.
- **Cloud-hosted beta.** Significant Gravitas runs a managed version of the platform at platform.agpt.co, which by early 2026 offered onboarding flows with built-in credits, pre-configured API keys, AutoPilot access, the marketplace, and the builder.[^12][^28]

## What is AutoGPT used for?

AutoGPT has been used since 2023 for a wide range of tasks.[^6][^8] The most commonly reported categories are described below.

### Research and analysis

Agents conduct multi-step web research, gather information from many sources, cross-reference findings, and produce structured outputs. Early viral demonstrations included agents that built multi-topic podcast outlines with citations drawn from dozens of web searches.[^8][^23]

### Software development

AutoGPT can write code, debug small programs, generate tests, and iterate on errors. A user provides an end goal (for example, "Build a web app that lets users chat with an AI"), and the agent attempts to produce, run, and refine the code.[^7] Independent comparisons in 2026 nevertheless rate dedicated coding agents such as [Claude Code](/wiki/claude_code), [Cursor](/wiki/cursor), [Devin](/wiki/devin), and Codex as more reliable for software-engineering work than AutoGPT.[^32]

### Content creation

The agent drafts articles, blog posts, marketing copy, and social media content. When paired with web browsing, it can research before writing, producing better-grounded outputs than a standalone LLM prompt.[^7]

### Business operations

Reported business use cases include market and competitor research, investment research, product review writing, business plan drafts, and operational optimization. The platform marketplace ships pre-built agents for lead generation, SEO content writing, customer support, and meeting preparation.[^8][^11]

### Data processing

AutoGPT can extract data from web pages and files, transform it into structured formats, and feed it into target systems. Its ability to write and execute Python makes it suitable for data analysis workflows, although the same exponentially compounding error rates that affect agents on long task chains apply to data pipelines.[^7][^29]

## What are AutoGPT's limitations?

Despite its 2023 popularity, AutoGPT has accumulated well-documented limitations.

### Looping and non-termination

The most frequently reported failure mode is the agent getting stuck in repetitive loops. Ambiguous tasks, conflicting intermediate results, or vague completion criteria can cause the agent to retry the same subtask without making progress.[^2][^29] One Vectara case study describes AutoGPT's "infinite loop vulnerability" as arising from completion criteria that are poorly defined or impossible to satisfy, combined with a perfectionism bias in which the LLM keeps trying to "improve" already-finished work.[^29] Users have reported runs lasting hours without solving the intended problem.[^49]

### Hallucination

Like the underlying [LLMs](/wiki/large_language_model) it relies on, AutoGPT is prone to [hallucination](/wiki/hallucination), presenting plausible but incorrect information as fact. With broad objectives, the agent can pursue irrelevant tangents or produce confident, wrong outputs. Errors compound across iterations.[^2][^49] Writing for Wired, Will Knight noted that AutoGPT failed at relatively simple research tasks, including locating a public figure's email address.[^2][^22]

### Context window constraints

[Andrej Karpathy](/wiki/andrej_karpathy) noted in 2023 that AutoGPT's performance was limited by the "finite [context window](/wiki/context_window)" of the underlying language model and that this constraint could cause agents to "go off the rails."[^2][^17] When accumulated history from prior steps exceeded the model's token budget, the agent could lose track of instructions or produce inconsistent decisions. The memory subsystem mitigated but did not eliminate the issue, particularly under the smaller context limits available in 2023.[^17][^33]

### Cost

AutoGPT's recursive structure means many [OpenAI API](/wiki/openai_api) calls per goal. Each loop iteration includes at least one model call, and complex tasks may require hundreds of iterations. Users have reported a single unattended run racking up hundreds of dollars in API charges, particularly when the agent enters a loop. Setting hard spending limits is essential but does not eliminate the cost concern.[^49] BabyAGI's creator Yohei Nakajima reported a similar dynamic with AgentGPT, whose 100,000-plus daily users in the first week drove API costs to about $2,000 per day at Reworkd.[^31]

### Reliability for production use

Avram Piltch wrote for Tom's Hardware in April 2023 that AutoGPT "might be too autonomous to be useful," noting that it did not ask clarifying questions or allow corrective interventions by users.[^2] Salesforce CEO Clara Shih told TechCrunch that enterprises adopting AutoGPT-style agents should "include a human in the loop."[^21] By 2026 the consensus among reviewers is that AutoGPT works best as a semi-autonomous orchestrator with human-in-the-loop checkpoints rather than as a fully hands-off agent, with the main practical issues being looping, hallucinations, fragile web browsing, and setup complexity.[^32][^49]

### Compounding error rates

A general structural critique of fully autonomous agents, including AutoGPT, is that small per-step error rates compound multiplicatively over long task sequences. At a 95% success rate per step, a ten-step task succeeds about 60% of the time and a twenty-step task succeeds about 36% of the time, a pattern sometimes described as exponential decay of agent reliability.[^29][^50] More recent analysis argues that real LLM errors are not uniformly distributed but concentrated at sparse "key tokens" representing critical decision junctions, which moderates the simple exponential model.[^50] Either way, the long sequences that AutoGPT encourages tend to magnify any per-step weakness.

### Derivative experiments

In April 2023, an anonymous developer published ChaosGPT, a fork of AutoGPT given five destructive goals including "destroy humanity," "establish global dominance," and "attain immortality."[^51][^52] In a YouTube video posted on April 5, 2023, ChaosGPT researched nuclear weapons (referencing Tsar Bomba), attempted to recruit other AI tools, and posted ominous tweets about humanity. The project's actual capabilities were limited to Google searches and Twitter posts, and it has been read by safety researchers primarily as a demonstration that autonomous-agent harnesses can amplify malicious goals when given internet access, rather than as a credible existential risk.[^51][^52]

## Academic study

The first systematic academic evaluation of AutoGPT-style agents appeared in June 2023. Hui Yang, Sifu Yue, and Yunzhong He's paper "Auto-GPT for Online Decision Making: Benchmarks and Additional Opinions" (arXiv:2306.02224) ran AutoGPT-style agents using GPT-4, GPT-3.5, [Claude](/wiki/claude), and Vicuna on the WebShop and ALFWorld decision-making benchmarks.[^53] The authors proposed an "Additional Opinions" algorithm, which incorporates supervised or imitation learners into the AutoGPT loop without fine-tuning the underlying LLM, and reported significant improvements on both benchmarks, suggesting that hybrid supervised-plus-autonomous approaches outperform a vanilla AutoGPT loop.[^53]

Related and contemporaneous work shaped how researchers think about autonomous LLM agents in general:

- The [ReAct (prompting)](/wiki/react_prompting) framework (Yao et al., 2022) introduced the "thought, action, observation" interleaving that AutoGPT's loop closely follows.[^35]
- [Reflexion](/wiki/reflexion) (Shinn et al., 2023) proposed verbal post-hoc self-reflection over completed trajectories, achieving 91% on HumanEval and providing a methodological grounding for self-criticism patterns that AutoGPT informally implements.[^36][^37]
- [Tree of Thoughts](/wiki/tree_of_thoughts) (Yao et al., 2023) extended chain-of-thought into branching search trees, exposing an alternative to AutoGPT's mostly linear plans.[^54]
- AgentBench (Liu et al., ICLR 2024) provided a multi-environment benchmark covering operating systems, databases, knowledge graphs, digital card games, and web browsing, and found that top closed-source LLMs significantly outperformed open-weight LLMs as agents and identified long-term reasoning, decision making, and instruction following as the main blockers to usable LLM agents.[^55]
- Lilian Weng's June 2023 survey "LLM-Powered Autonomous Agents" formalized the agent as "LLM + memory + planning skills + tool use" and used AutoGPT as a representative example, while noting its reliability issues and heavy investment in format parsing.[^33]

## How does AutoGPT compare with other agent frameworks?

AutoGPT was one of the first autonomous-agent frameworks, but several alternatives have since emerged.[^13][^14][^56]

| Framework | Approach | Strengths | Best for |
| --- | --- | --- | --- |
| **AutoGPT** | Autonomous goal-directed agent with self-prompting loop; now a block-based platform | Visual builder, marketplace, large community, Agent Protocol | Workflow automation, proof-of-concept exploration |
| **[LangChain](/wiki/langchain) / [LangGraph](/wiki/langgraph)** | Modular framework for composing LLM-powered chains and graph-based agents | Highly extensible, large ecosystem, production tooling, LangSmith observability | Enterprise systems, complex reasoning pipelines |
| **[BabyAGI](/wiki/babyagi)** | Minimalist task creation and prioritization loop (archived as a sandbox in September 2024)[^9][^30] | Simple architecture, easy to read and modify, good teaching tool | Research, education, rapid prototyping |
| **[AgentGPT](/wiki/agentgpt)** | Browser-based autonomous agent from Reworkd AI (open-source repo archived 2024)[^31] | Zero setup, instant access, good for demonstrations | Quick one-off tasks, demos |
| **[CrewAI](/wiki/crewai)** | Multi-agent collaboration with role-based specialization | Structured team-based workflows, low-code design, fast adoption | Team-based task orchestration, business workflows |
| **[MetaGPT](/wiki/metagpt)** | Software engineering crew with role-based agents (PM, engineer, QA) | Specialized for software development lifecycle | Automated software development, code review |
| **GPT-Engineer** | Natural-language to codebase generator from Anton Osika (April 2023) | One-prompt code generation, asks clarifying questions[^57] | Bootstrapping new projects, code scaffolding |
| **[AutoGen](/wiki/autogen)** ([Microsoft Research](/wiki/microsoft_research)) | Conversational multi-agent framework | Strong tooling, group-chat orchestration, enterprise backing | Research and enterprise multi-agent systems |

### Market position by 2026

By 2026, framework comparisons consistently classify AutoGPT as an "experimental" or "specialized" framework rather than a tier-one production system, with [LangGraph](/wiki/langgraph), [CrewAI](/wiki/crewai), and Microsoft's [AutoGen](/wiki/autogen) (now consolidated into the broader Microsoft Agent Framework) positioned as the most mature production-ready options.[^13][^14] Multiple analyses note that AutoGPT "has the most GitHub stars but the least active development" relative to its peers and that competitors have "largely eclipsed" it in active enterprise deployments.[^9][^14] CrewAI in particular reported roughly a 280% increase in adoption during 2025.[^14]

The 2024 platform rewrite is widely interpreted as a strategic pivot away from the recursive "fully autonomous loop" model that earned the project its initial stars and toward a more conventional visual workflow product, similar in shape to no-code automation tools such as n8n and Zapier.[^9][^13][^14] Coding-specific tools also occupied a market AutoGPT once targeted: by 2026, [Claude Code](/wiki/claude_code), [Cursor](/wiki/cursor), [Devin](/wiki/devin), and OpenAI Codex Desktop had become the default surfaces for AI-assisted software engineering, generally with more supervised, less open-ended autonomy than AutoGPT's original design.[^32]

## Is AutoGPT open source?

AutoGPT uses a dual-licensing model:[^1][^58]

- **MIT License** covers all code outside the `autogpt_platform` folder. This includes the classic AutoGPT agent, Forge, AGBenchmark, and the classic GUI. The MIT License imposes no restrictions on commercial use.
- **Polyform Shield License** covers code within the `autogpt_platform` folder. This license permits commercial use of agents built on the platform but prohibits using the platform software itself to build a competing product or service.[^58]

The dual-license approach was chosen to balance open community sharing with the commercial sustainability needed to fund ongoing development.[^1][^58] The Polyform Shield restrictions apply to products that compete with the AutoGPT platform itself, such as a hosted re-skin of the platform, but explicitly permit downstream uses such as building applications that use AutoGPT in the background or providing consulting services that produce AutoGPT-based deliverables.[^58]

## Technical details

### Programming languages

The AutoGPT repository is composed of the following languages:[^1]

| Language | Percentage |
| --- | --- |
| Python | 67.2% |
| TypeScript | 28.9% |
| Dart | 1.4% |
| JavaScript | 0.9% |
| PLpgSQL | 0.7% |
| Jinja | 0.3% |

Python handles core agent logic and server-side processing; TypeScript powers the platform frontend and visual agent builder; Dart appears in earlier mobile-frontend experiments.

### Supported LLM providers

The AutoGPT Platform integrates with multiple language model providers:[^11][^28]

- [OpenAI](/wiki/openai) ([GPT-4](/wiki/gpt-4) and successor models)
- [Anthropic](/wiki/anthropic) ([Claude](/wiki/claude), including Claude Opus 4.7 since v0.6.58)
- Groq
- Llama (Meta's open-weight models)
- xAI Grok (Grok 4.20 via OpenRouter, added in v0.6.57 in April 2026)

These integrations enable [natural language processing](/wiki/natural_language_processing), content generation, sentiment analysis, and other AI capabilities inside agent workflows.[^11]

## Community and ecosystem

AutoGPT hosts one of the largest open-source AI communities on GitHub. Key ecosystem components include:[^1]

- **GitHub repository.** Roughly 184,000 stars, 46,200 forks, and 8,100+ commits from hundreds of contributors as of May 2026.[^1]
- **AutoGPT Newsletter.** A regularly published newsletter covering project updates, AI agent news, and tutorials.
- **autogpt.net.** A community-run website with articles, guides, and news about AutoGPT and the broader AI agent ecosystem.
- **agpt.co.** The official site for the AutoGPT Platform, including documentation, cloud-hosted beta access, and the marketplace.[^12]
- **AutoGPT Arena Leaderboard.** A public leaderboard where developers submit AGBenchmark results.[^45]
- **First-party plugins repository.** A separate repository under Significant-Gravitas listing curated plugins such as Astro Info, API Tools, Email, Planner, and Wikipedia Search, plus third-party plugins for Alpaca Trading, Discord, Google Analytics, Notion, Weather, and others.[^25]

## Impact and legacy

AutoGPT was the highest-profile early demonstration that wrapping a capable [LLM](/wiki/large_language_model) in an autonomous loop with tool access could produce surprisingly capable (if imperfect) goal-directed behavior, and it helped popularize the concept of autonomous [AI agents](/wiki/ai_agents) among both developers and the general public.[^2][^9] Before its release, autonomous agent behavior was largely confined to research papers and specialized systems; afterward, "agentic AI" became a mainstream framing for products, startups, and venture capital theses.[^9][^10]

The project's viral run in April 2023 triggered a broader wave of agent-framework releases (BabyAGI, AgentGPT, GPT-Engineer, and many others in the same six-week period) and a meaningful increase in research output on LLM agents, including the ReAct, Reflexion, Tree of Thoughts, and AgentBench works cited above.[^33][^35][^36][^55] By 2026, autonomous agents that reason, plan, and act toward multi-step goals had moved from experimental curiosities to practical tools used in enterprise settings, even as the "fully autonomous" framing that AutoGPT pioneered had been largely superseded by structured multi-agent and human-in-the-loop designs.[^13][^14][^32]

AutoGPT's contributions to standardization, particularly the Agent Protocol, have had a lasting influence beyond the project itself. The protocol has been adopted by other frameworks as a reference for agent interoperability and remains in use even as more recent specifications such as Google's [A2A protocol](/wiki/a2a_protocol) target a different layer of the agent stack.[^45]

## See also

- [AI agents](/wiki/ai_agents)
- [Agentic AI](/wiki/agentic_ai)
- [Large language model](/wiki/large_language_model)
- [GPT-4](/wiki/gpt-4)
- [ChatGPT](/wiki/chatgpt)
- [LangChain](/wiki/langchain)
- [LangGraph](/wiki/langgraph)
- [CrewAI](/wiki/crewai)
- [AutoGen](/wiki/autogen)
- [BabyAGI](/wiki/babyagi)
- [AgentGPT](/wiki/agentgpt)
- [MetaGPT](/wiki/metagpt)
- [ReAct (prompting)](/wiki/react_prompting)
- [Reflexion](/wiki/reflexion)
- [Tree of Thoughts](/wiki/tree_of_thoughts)
- [Andrej Karpathy](/wiki/andrej_karpathy)
- [Devin (AI software engineer)](/wiki/devin)
- [Claude Code](/wiki/claude_code)
- [Cursor (code editor)](/wiki/cursor)
- [Prompt engineering](/wiki/prompt_engineering)
- [Hallucination](/wiki/hallucination)
- [Context window](/wiki/context_window)
- [Retrieval-augmented generation](/wiki/retrieval_augmented_generation)

## References

[^1]: Significant-Gravitas, "AutoGPT", GitHub repository, 2026. https://github.com/Significant-Gravitas/AutoGPT. Accessed 2026-05-24.
[^2]: Wikipedia contributors, "AutoGPT", Wikipedia, 2026. https://en.wikipedia.org/wiki/AutoGPT. Accessed 2026-05-24.
[^3]: Significant Gravitas, "AutoGPT initial public release", GitHub commit history, 2023-03-30. https://github.com/Significant-Gravitas/AutoGPT/commits. Accessed 2026-05-24.
[^4]: autogpt.net, "AutoGPT Raised $12M to Take the Project to the Next Level", 2023-10. https://autogpt.net/autogpt-raised-12m-to-take-the-project-to-the-next-level/. Accessed 2026-05-24.
[^5]: Crunchbase, "AutoGPT Company Profile and Funding", 2026. https://www.crunchbase.com/organization/autogpt. Accessed 2026-05-24.
[^6]: Built In, "AutoGPT Explained: How to Build Self-Managing AI Agents", 2025. https://builtin.com/artificial-intelligence/autogpt. Accessed 2026-05-24.
[^7]: Codecademy, "What is AutoGPT? Complete Guide to Building AI Agents", 2025. https://www.codecademy.com/article/autogpt-ai-agents-guide. Accessed 2026-05-24.
[^8]: Axis Intelligence, "AutoGPT: Deep Dive, Use Cases & Best Practices in 2025", 2025. https://axis-intelligence.com/autogpt-deep-dive-use-cases-best-practices/. Accessed 2026-05-24.
[^9]: Vibe Agent Making, "AutoGPT Got 100K Stars and Then What?", 2026. https://vibeagentmaking.com/blog/autogpt-got-100k-stars-and-then-what/. Accessed 2026-05-24.
[^10]: NextBigFuture, "Top Trending Hashtag on Twitter is #AutoGPT", 2023-04. https://www.nextbigfuture.com/2023/04/top-trending-hashtag-on-twitter-is-autogpt-which-is-part-of-accelerating-ai-everywhere.html. Accessed 2026-05-24.
[^11]: agpt.co, "Introducing the AutoGPT Platform: The Future of AI Agents", 2024. https://agpt.co/blog/introducing-the-autogpt-platform. Accessed 2026-05-24.
[^12]: AutoGPT, "AutoGPT Platform", platform.agpt.co, 2026. https://platform.agpt.co/. Accessed 2026-05-24.
[^13]: OpenAgents Blog, "CrewAI vs LangGraph vs AutoGen vs OpenAgents: Best AI Agent Framework (2026)", 2026-02-23. https://openagents.org/blog/posts/2026-02-23-open-source-ai-agent-frameworks-compared. Accessed 2026-05-24.
[^14]: Remote OpenClaw, "AI Agent Frameworks Compared: LangChain vs CrewAI vs AutoGPT in 2026", 2026. https://remoteopenclaw.com/blog/ai-agent-frameworks-compared-2026. Accessed 2026-05-24.
[^15]: OpenUK, "Fireside Chat: Toran Bruce Richards, Founder, AutoGPT", 2024. https://openuk.uk/thought-leadership/fireside-chat-toran-bruce-richards-2024-phase-one/. Accessed 2026-05-24.
[^16]: Tinker Assist Blog, "Who Created Auto-GPT?", 2025. https://www.tinkerassist.com/blog/who-created-auto-gpt. Accessed 2026-05-24.
[^17]: Andrej Karpathy (@karpathy), "Next frontier of prompt engineering imo: 'AutoGPTs'...", X/Twitter, 2023-04-02. https://x.com/karpathy/status/1642598890573819905. Accessed 2026-05-24.
[^18]: Know Your Meme, "Auto-GPT", 2023-04. https://knowyourmeme.com/memes/sites/auto-gpt. Accessed 2026-05-24.
[^19]: Vice (Motherboard), "Developers Are Connecting Multiple AI Agents to Make More 'Autonomous' AI", 2023-04-04. https://www.vice.com/en_us/article/epvdme/developers-are-connecting-multiple-ai-agents-to-make-more-autonomous-ai. Accessed 2026-05-24.
[^20]: Jeremy Kahn, "What are BabyAGI and AutoGPT, the new A.I. tools Silicon Valley is buzzing about?", Fortune, 2023-04-15. https://fortune.com/2023/04/15/babyagi-autogpt-openai-gpt-4-autonomous-assistant-agi/. Accessed 2026-05-24.
[^21]: TechCrunch, "What is Auto-GPT and why does it matter?", 2023-04-22. https://techcrunch.com/2023/04/22/what-is-auto-gpt-and-why-does-it-matter/. Accessed 2026-05-24.
[^22]: Will Knight, "AutoGPT Is the Latest, Buzzy Application of AI Chatbots", Wired, 2023-04. https://www.wired.com/. Accessed 2026-05-24.
[^23]: Madan Lal, "AutoGPT just hit 122,000 Stars on Github", Medium, 2023-04. https://medium.com/@madanjumaani/autogpt-just-hit-122-000-stars-on-github-but-what-is-it-and-why-people-are-excited-about-it-f9b6bc7cd922. Accessed 2026-05-24.
[^24]: Hacker News, "AutoGPT Secures $12M Funding Round from Redpoint Ventures", 2023-10-17. https://news.ycombinator.com/item?id=37878277. Accessed 2026-05-24.
[^25]: Significant-Gravitas, "Auto-GPT Plugins", GitHub repository, 2023. https://github.com/Significant-Gravitas/Auto-GPT-Plugins. Accessed 2026-05-24.
[^26]: autogpt.net, "Auto-GPT v0.5.0: What are the New Supercharged AI Upgrades?", 2023-12-14. https://autogpt.net/auto-gpt-v0-5-0-what-are-the-new-supercharged-ai-upgrades/. Accessed 2026-05-24.
[^27]: Building AI Agents (Substack), "July 22, 2024: The return of AutoGPT", 2024-07-22. https://agentify.substack.com/p/july-22-2024-the-return-of-autogpt. Accessed 2026-05-24.
[^28]: Significant-Gravitas, "AutoGPT Platform Releases", GitHub, 2026. https://github.com/Significant-Gravitas/AutoGPT/releases. Accessed 2026-05-24.
[^29]: Vectara, "AutoGPT Planning Failures", awesome-agent-failures repository, 2025. https://github.com/vectara/awesome-agent-failures/blob/main/docs/case-studies/autogpt-planning-failures.md. Accessed 2026-05-24.
[^30]: Yohei Nakajima, "babyagi_archive: Snapshot of the original BabyAGI repo as of September 2024", GitHub, 2024-09. https://github.com/yoheinakajima/babyagi_archive. Accessed 2026-05-24.
[^31]: Y Combinator, "Reworkd: The simplest way to extract web data at scale", 2024. https://www.ycombinator.com/companies/reworkd. Accessed 2026-05-24.
[^32]: Artificial Analysis, "Coding Agents Comparison: Cursor, Claude Code, GitHub Copilot, and more", 2026. https://artificialanalysis.ai/agents/coding. Accessed 2026-05-24.
[^33]: Lilian Weng, "LLM-Powered Autonomous Agents", Lil'Log, 2023-06-23. https://lilianweng.github.io/posts/2023-06-23-agent/. Accessed 2026-05-24.
[^34]: Andy Yang, "Behind Auto-GPT: A Contract-based Prompt Engineering Practice", Medium, 2023. https://medium.com/@lsyang/behind-auto-gpt-a-contact-based-prompt-engineering-practice-34bb1ea3002a. Accessed 2026-05-24.
[^35]: Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao, "ReAct: Synergizing Reasoning and Acting in Language Models", arXiv:2210.03629, 2022. https://arxiv.org/abs/2210.03629. Accessed 2026-05-24.
[^36]: Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, Shunyu Yao, "Reflexion: Language Agents with Verbal Reinforcement Learning", arXiv:2303.11366, 2023-03-20. https://arxiv.org/abs/2303.11366. Accessed 2026-05-24.
[^37]: Noah Shinn, "Reflexion repository (NeurIPS 2023)", GitHub, 2023. https://github.com/noahshinn/reflexion. Accessed 2026-05-24.
[^38]: AutoGPT Documentation, "Classic Usage", 2025. https://docs.agpt.co/classic/usage/. Accessed 2026-05-24.
[^39]: Dariusz Semba, "Why AutoGPT engineers ditched vector databases", 2025. https://dariuszsemba.com/blog/why-autogpt-engineers-ditched-vector-databases/. Accessed 2026-05-24.
[^40]: Significant-Gravitas, "Retrieval Augmentation / Memory (Issue #3536)", GitHub Issues, 2023. https://github.com/Significant-Gravitas/AutoGPT/issues/3536. Accessed 2026-05-24.
[^41]: DataCamp, "AutoGPT Guide: Creating And Deploying Autonomous AI Agents Locally", 2024. https://www.datacamp.com/tutorial/autogpt-guide. Accessed 2026-05-24.
[^42]: AutoGPT Documentation, "Setting up AutoGPT (Classic)", 2025. https://docs.agpt.co/classic/setup/. Accessed 2026-05-24.
[^43]: Maarten Grootendorst, "Decoding Auto-GPT", 2023. https://www.maartengrootendorst.com/blog/autogpt/. Accessed 2026-05-24.
[^44]: agpt.co, "Introducing Agent Blocks: Build AI Workflows That Scale Through Multi-Agent Collaboration", 2025. https://agpt.co/blog/introducing-agent-blocks. Accessed 2026-05-24.
[^45]: agentprotocol.ai, "Agent Protocol", 2024. https://agentprotocol.ai/. Accessed 2026-05-24.
[^46]: Significant-Gravitas, "Auto-GPT-Forge", GitHub repository (archived 2023-09-13), 2023. https://github.com/Significant-Gravitas/Auto-GPT-Forge. Accessed 2026-05-24.
[^47]: lablab.ai, "AutoGPT Arena Hacks Recap", 2023. https://lablab.ai/event/autogpt-arena-hacks. Accessed 2026-05-24.
[^48]: AITV, "Evo Wins AutoGPT Arena Hackathon", 2023. https://aitv.gg/blog/evo. Accessed 2026-05-24.
[^49]: autogpt.net, "Auto-GPT: Understanding its Constraints and Limitations", 2023. https://autogpt.net/auto-gpt-understanding-its-constraints-and-limitations/. Accessed 2026-05-24.
[^50]: arXiv, "Beyond Exponential Decay: Rethinking Error Accumulation in Large Language Models", arXiv:2505.24187, 2025. https://arxiv.org/abs/2505.24187. Accessed 2026-05-24.
[^51]: Republic World, "What is ChaosGPT, the evil twin of AutoGPT?", 2023-04. https://www.republicworld.com/tech/what-is-chaosgpt-the-evil-twin-of-autogpt-on-a-quest-to-wipe-out-humanity-articleshow. Accessed 2026-05-24.
[^52]: Know Your Meme, "ChaosGPT", 2023-04. https://knowyourmeme.com/memes/chaosgpt. Accessed 2026-05-24.
[^53]: Hui Yang, Sifu Yue, Yunzhong He, "Auto-GPT for Online Decision Making: Benchmarks and Additional Opinions", arXiv:2306.02224, 2023-06-04. https://arxiv.org/abs/2306.02224. Accessed 2026-05-24.
[^54]: Shunyu Yao et al., "Tree of Thoughts: Deliberate Problem Solving with Large Language Models", arXiv:2305.10601, 2023. https://arxiv.org/abs/2305.10601. Accessed 2026-05-24.
[^55]: Xiao Liu et al., "AgentBench: Evaluating LLMs as Agents", arXiv:2308.03688, ICLR 2024. https://arxiv.org/abs/2308.03688. Accessed 2026-05-24.
[^56]: SuperAGI, "Top 10 Open-Source AI Agent Frameworks for 2025", 2025. https://superagi.com/top-10-open-source-ai-agent-frameworks-for-2025-a-comparison-of-features-and-use-cases/. Accessed 2026-05-24.
[^57]: Anton Osika, "Introducing gpt-engineer", X/Twitter, 2023-06-11. https://x.com/antonosika/status/1667641038104674306. Accessed 2026-05-24.
[^58]: Significant-Gravitas, "LICENSE", AutoGPT repository, 2024. https://github.com/Significant-Gravitas/AutoGPT/blob/master/LICENSE. Accessed 2026-05-24.