AutoGPT Agent
Last reviewed
Apr 28, 2026
Sources
12 citations
Review status
Source-backed
Revision
v6 · 4,870 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Last reviewed
Apr 28, 2026
Sources
12 citations
Review status
Source-backed
Revision
v6 · 4,870 words
Add missing citations, update stale details, or suggest a clearer explanation.
| AutoGPT Agent | |
|---|---|
![]() | |
| Information | |
| Name | AutoGPT Agent |
| Platform | ChatGPT |
| Store | GPT Store |
| Model | GPT-4 |
| Category | Productivity |
| Description | Your personal AI agent will plan, research, strategize and work to complete tasks semi-autonomously using multi-modal tools as needed. Complete tasks with just a few keystrokes. v1.3 |
| Developer | agilayer.com |
| OpenAI URL | https://chat.openai.com/g/g-POb5UhhJ6-autogpt-agent |
| Chats | 7,000 |
| Actions | Yes |
| Web Browsing | Yes |
| DALL·E Image Generation | Yes |
| Code Interpreter | Yes |
| Free | Yes |
| Available | Yes |
| Updated | 2024-01-24 |
AutoGPT Agent is a Custom GPT for ChatGPT listed in the GPT Store, built by agilayer.com. The GPT borrows its name and concept from AutoGPT, the open source autonomous agent project released by Toran Bruce Richards in March 2023, and packages a similar promise into a single ChatGPT system prompt. It plans, researches, and works through tasks semi-autonomously by chaining together the web browsing, code interpreter, and DALL·E tools that ship with ChatGPT, plus a Zapier webhook action for sending data out to external services [1][2].
The broader concept the Custom GPT is named after, namely an LLM that loops on its own output until a goal is met, was one of the defining ideas of the 2023 generative AI boom. AutoGPT itself peaked at well over 100,000 GitHub stars within weeks of launch and became, at the time, the fastest growing repository in GitHub history. It went on to inspire a wave of imitators, including BabyAGI, AgentGPT, GodMode, and ChaosGPT, and reshaped how the wider community thought about AI agents [2][3][4].
This article covers both: the specific AutoGPT Agent Custom GPT in the GPT Store, and the original AutoGPT project that gave it its name. The two are not affiliated; the Custom GPT is a third party application that adopts the AutoGPT label.
AutoGPT Agent (the Custom GPT) is a GPT-4 based assistant published on the GPT Store by agilayer.com. As of the last verified update on January 24, 2024, the listing reported around 7,000 chats and version 1.3 of the agent. It is free to use for any ChatGPT Plus subscriber with access to the store [1].
The public GPT Store listing surfaces four suggested conversation starters that hint at the intended use cases:
These cover four practical workflows that combine ChatGPT's built in tools: strategy and copy generation, image creation through DALL·E, web research with sentiment analysis, and data visualization through the code interpreter. The starters are deliberately broad, in the spirit of a generalist agent rather than a narrow assistant [1].
The AutoGPT Agent Custom GPT enables every standard ChatGPT tool plus a custom action. The combination is what allows the assistant to behave in a quasi autonomous fashion within a single conversation.
| Capability | Provided by | Notes |
|---|---|---|
| Reasoning and planning | GPT-4 | The base model handles task decomposition through chain of thought style prompting. |
| Web browsing | ChatGPT browsing tool | Fetches live web pages to gather facts, news, and competitive research. |
| Code interpreter | ChatGPT Python sandbox | Runs Python in an isolated environment to crunch data, build charts, and process files. |
| Image generation | DALL·E 3 | Produces logos, marketing visuals, and other graphics inline. |
| External actions | Zapier webhook | Sends structured JSON payloads to a configured Zapier hook for downstream automation, including newsletter signup and notifications. |
The Zapier integration is the only first party custom action exposed by the GPT. According to its OpenAPI 3.1 specification, it posts an email field to a fixed Zapier webhook URL at hooks.zapier.com/hooks/catch/14137578/3kabw2k/. The action uses no authentication, which limits how sensitive the payload should be in practice [1].
The AutoGPT Agent Custom GPT is not part of the Significant Gravitas AutoGPT project and shares no code with it. It is a separate product that adopts the AutoGPT branding because the name became shorthand, throughout 2023, for the idea of a goal seeking LLM agent. The next sections cover that original project in detail, since most users searching for AutoGPT in the GPT Store are reaching for that broader idea.
AutoGPT is an open source autonomous agent framework written in Python and TypeScript that wraps OpenAI's GPT-4 (and later models) in a planning, action, and reflection loop. The user provides a high level goal in natural language, and the agent decomposes that goal into smaller subtasks, executes each one through tools such as web browsing, file I/O, and code execution, evaluates the results, and iterates until the objective is met or a stop condition fires [2][3].
Released on March 30, 2023 by Toran Bruce Richards under the name Auto-GPT, the project quickly became one of the most discussed pieces of software of the year. Within roughly two weeks of launch it crossed 30,000 GitHub stars, and within months it had passed 100,000, briefly making it the most popular AI repository on the platform. As of April 2026, the Significant-Gravitas/AutoGPT repository sits at around 184,000 stars and 46,200 forks, with active development now centered on the AutoGPT Platform rather than the original command line agent [2][5].
AutoGPT was created by Toran Bruce Richards, a 2020 graduate of Edinburgh Napier University with a Bachelor of Science in Game Development. Richards is the founder of Significant Gravitas Ltd., a UK based company originally focused on applying software techniques from the video game industry to non gaming use cases. On X (formerly Twitter) he goes by the handle @SigGravitas [2][6].
The initial Auto-GPT release landed on March 30, 2023, just two weeks after OpenAI launched GPT-4. Richards has said the goal was to build an AI system that could pursue objectives with a long term outlook and respond to real time feedback, without the operator having to hand hold each step. He chose the open source route in part because he was uncomfortable with the idea of locking such a tool behind a single corporate API [2].
The timing was almost ideal. ChatGPT had pushed generative AI into the mainstream a few months earlier, and developers were hungry for ways to combine LLMs with tools, memory, and the open web. Auto-GPT made that idea concrete in a few hundred lines of Python that anyone could clone and run [2][3].
The project's growth curve in April 2023 was unusually steep, even by the standards of AI hype that year:
| Date | Milestone |
|---|---|
| March 30, 2023 | Initial Auto-GPT release on GitHub |
| April 3, 2023 | Top trending repository on GitHub |
| April 4, 2023 | Motherboard publishes the first major mainstream article |
| April 12, 2023 | Crosses 30,000 GitHub stars |
| Mid April 2023 | Top trending hashtag on Twitter; ChaosGPT goes viral |
| Mid 2023 | Crosses 100,000 GitHub stars |
| October 2023 | Significant Gravitas raises $12M Series A |
Major outlets including TechCrunch, Fortune, The Verge, and Fast Company covered the project. Demonstrations on Twitter and YouTube showed Auto-GPT trying to plan a podcast, build a small web application end to end, conduct market research, or compose a business plan. Many of the demos failed in entertaining ways, but the failures themselves drove engagement [2][4][7].
In October 2023, Significant Gravitas Ltd. announced a $12 million venture funding round. The round was led by Redpoint Ventures, with participation from GitHub through GitHub Ventures. The capital was earmarked for further development of the AutoGPT platform and for hiring around the open source project. The company remained registered in the United Kingdom despite taking U.S. capital [2][8].
The original Auto-GPT agent is built around a tight planning and execution loop, supported by short and long term memory and a small set of tool integrations. The exact details have shifted across versions, but the overall shape has been stable since the first major releases.
On each iteration, AutoGPT prompts the underlying LLM with the user goal, the relevant memory, the tools available, and a constrained output schema. The model is expected to return three things: a plan, a justification, and a single action command. The agent code parses the action, runs the corresponding tool (open URL, read file, run shell command, generate image, and so on), feeds the result back into context, and loops [3][9].
| Step | Description |
|---|---|
| 1. Goal definition | User states a high level objective in plain English. |
| 2. Plan | The model writes a plan and decomposes the goal into the next subtask. |
| 3. Reasoning | The model explains why this subtask makes sense given progress so far. |
| 4. Action | The model emits exactly one structured action command (for example, web_search, read_file, or execute_python_file). |
| 5. Tool execution | The agent runs the action and captures stdout, stderr, and any return value. |
| 6. Memory write | The result is appended to short term memory, with embeddings optionally pushed into long term storage. |
| 7. Self review | The model evaluates progress and either continues, replans, or signals completion. |
The loop continues until the model emits a finish action, the user halts execution, or a hard cap on iterations or token spend is hit. In practice the iteration cap is what saves most users from runaway costs [3][9][10].
AutoGPT distinguishes between short term and long term memory. Short term memory is the current message buffer, capped by the model context window. Earlier versions kept around the first nine messages plus the most recent tool outputs to stay inside GPT-4's 8K or 32K context [9].
Long term memory is a key value store of (vector, text) pairs. To embed text, AutoGPT initially used OpenAI's text-embedding-ada-002 API. The vectors and their source text were stored in a vector database, with the agent running a top K nearest neighbour search over them whenever it needed to recall older context [9][10].
| Memory tier | Storage | Purpose |
|---|---|---|
| Working context | In-process message buffer | Recent reasoning and tool outputs |
| Short term cache | Local files | Stable session state across loop iterations |
| Long term store | Vector database (historical) or local datastore (current) | Semantic recall of older facts and observations |
In early 2023, AutoGPT 0.2.x supported Pinecone, Milvus, Redis, and Weaviate as long term memory backends. By late 2023, the maintainers controversially dropped all third party vector databases and shifted to a simpler local datastore in order to reduce dependency churn and configuration complexity. This decision drew pushback from parts of the community that valued the flexibility, but it made the agent easier to install and run for newcomers [9][10].
The classic AutoGPT agent ships with a small but expressive toolkit:
| Tool | Capability |
|---|---|
| Web search | Calls a search engine (originally Google through SerpAPI) to find sources |
| Web browsing | Fetches a URL and extracts the readable text or specific elements |
| File I/O | Reads, writes, and lists files in a sandboxed workspace/ directory |
| Code execution | Runs Python files in a Docker container or local sandbox |
| Shell commands | Issues whitelisted shell commands when enabled |
| Image generation | Generates images through DALL·E or Stable Diffusion |
| Speech | Optionally speaks reasoning aloud through ElevenLabs or local TTS |
| Plugins | Loads community plugins for things like Twitter, Telegram, or LangChain integrations |
The plugin architecture, introduced in mid 2023, was a major contributor to AutoGPT's flexibility. Anyone could write a Python class implementing a small interface and drop it into the plugins/ folder. The agent loaded plugins at startup and exposed their actions to the model alongside the built in commands [3][10].
AutoGPT's release history splits cleanly into two eras: the classic command line agent, and the AutoGPT Platform.
| Version | Date | Notes |
|---|---|---|
| v0.1.0 | March 30, 2023 | Initial public release of Auto-GPT |
| v0.2.x | April to June 2023 | Vector database memory, plugins, multi LLM support |
| v0.3.x | Mid 2023 | Improved planning, broader plugin ecosystem |
| v0.4.x | Late 2023 | Re-Architecture, Forge toolkit, agent benchmarking |
| v0.5.0 | December 2023 | Major refactor of the classic agent for stability |
| v0.5.1 | April 2024 | Final maintenance release of classic AutoGPT |
| Platform Beta v0.4.0 | December 2024 | First broad release of the new AutoGPT Platform with low code builder and marketplace |
| Platform Beta v0.6.x | 2025 to 2026 | Cloud hosting, blocks system, multi agent support, expanded marketplace |
| Platform Beta v0.6.57 | April 22, 2026 | Latest release as of this writing |
From v0.5.x onward, the project's centre of gravity moved to the platform. The classic command line agent still ships with the repository for users who want the original experience, but it no longer receives major feature work [5][8].
Distilling the project across versions, AutoGPT offers a recognisable feature set that has shaped how subsequent agent frameworks pitch themselves:
| Feature | Description |
|---|---|
| Goal directed autonomy | Operates from a single high level objective rather than turn by turn instructions |
| Self prompting loop | Generates its own subprompts and decides which tool to invoke next |
| Tool use | First class web browsing, file I/O, code execution, and image generation |
| Memory | Short term context plus optional long term embeddings store |
| Plugin system | Community plugins for Twitter, Telegram, search providers, and more |
| Multi model support | OpenAI GPT-3.5 and GPT-4 by default, with optional support for Anthropic, Google, and local models |
| Sandboxed execution | Code runs inside Docker containers by default to limit blast radius |
| Logging and replay | Detailed logs of each loop iteration for debugging and analysis |
| Benchmarking | The Forge toolkit and agbenchmark let developers compare agent performance |
| Cloud platform | Newer AutoGPT Platform offers hosted agents, marketplace, and a low code builder |
Significant Gravitas Ltd. is the UK based company behind AutoGPT. It was founded by Toran Bruce Richards before AutoGPT existed and originally positioned itself around bringing video game style production techniques to other industries. After Auto-GPT went viral in April 2023, the company pivoted into a full time AI focus [2].
| Detail | Value |
|---|---|
| Founder | Toran Bruce Richards |
| Headquarters | United Kingdom |
| Founded | Pre 2023 (under prior focus); pivoted to AI in 2023 |
| Flagship product | AutoGPT and the AutoGPT Platform |
| Funding | $12 million Series A in October 2023 |
| Lead investor | Redpoint Ventures |
| Other investor | GitHub Ventures |
| Repository | github.com/Significant-Gravitas/AutoGPT |
| Website | agpt.co |
The Redpoint led round, announced in October 2023, was unusual for a project that began as a side project rather than a structured startup. Redpoint partner Erica Brescia previously ran GitHub as COO, which likely smoothed the introduction. GitHub itself joined the round through GitHub Ventures, a fitting choice given how much of AutoGPT's distribution flowed through GitHub stars and forks [2][8].
AutoGPT was first to popularise the idea of a goal seeking LLM agent, but the space filled in quickly. By late 2023, LangChain, LlamaIndex, AutoGen, and a stream of forks were competing for developer mindshare.
| Framework | Origin | Primary focus | Language | Multi agent | Maintainer |
|---|---|---|---|---|---|
| AutoGPT | March 2023 | Single autonomous agent with tools and memory | Python, TypeScript | Limited (via platform blocks) | Significant Gravitas |
| BabyAGI | April 2023 | Minimalist task list driven agent | Python | No | Yohei Nakajima |
| AgentGPT | April 2023 | Browser hosted clone of AutoGPT | TypeScript | No | Reworkd |
| GodMode | April 2023 | Desktop wrapper around AutoGPT and AgentGPT | TypeScript | No | Lalit Kale |
| ChaosGPT | April 2023 | Demonstrative misuse fork with destructive prompts | Python | No | Anonymous |
| LangChain | October 2022 | General LLM application framework with agents and tools | Python, JS | Through extensions | LangChain Inc. |
| LlamaIndex | November 2022 | Retrieval and indexing for LLMs | Python | Through extensions | LlamaIndex Inc. |
| AutoGen | September 2023 | Multi agent conversation framework | Python, .NET | Yes (group chat) | Microsoft Research |
| CrewAI | 2023 | Role based multi agent crews | Python | Yes (crew abstraction) | CrewAI Inc. |
| LangGraph | 2024 | Graph based stateful agent workflows | Python, JS | Yes | LangChain Inc. |
| GPT-4 Code Interpreter | March 2023 | OpenAI hosted Python sandbox in ChatGPT | Closed | No | OpenAI |
| ChatGPT plugins | March 2023 | OpenAI's first party tool use system | Closed | No | OpenAI |
A few practical points stand out. AutoGPT favours a single agent doing many things, while AutoGPT and CrewAI favour multiple specialised agents talking to each other. LangChain is much broader: it provides building blocks (chains, agents, retrievers, callbacks) rather than a single ready to run agent. LlamaIndex started as a retrieval focused library and slowly grew agent capabilities. ChatGPT plugins and Code Interpreter were OpenAI's parallel attempts at the same problem, with much tighter integration but no open source story [3][7][11].
For most early adopters, the choice in 2023 looked roughly like this: AutoGPT for raw autonomy and tool breadth, BabyAGI for clean task lists and stable loops, AgentGPT or GodMode for a usable UI, and LangChain for building your own agent from primitives. Many teams ended up combining ideas from several of them [4][7].
AutoGPT's reception was unusually polarised. On one side were developers and futurists who argued that the project hinted at the first real glimpse of artificial general intelligence at consumer scale. On the other side were skeptics who pointed out that most demonstrations leaked, looped, or hallucinated their way to nowhere [3][4][7].
The GitHub history is the cleanest measurable proxy for impact:
| Metric | Value |
|---|---|
| Stars at peak (mid 2023) | Over 150,000, climbing past 180,000 by 2024 |
| Stars (April 2026) | Approximately 184,000 |
| Forks (April 2026) | Approximately 46,200 |
| Programming languages | Python (~70%), TypeScript (~28%) |
| License | MIT for the classic agent; Polyform Shield for the AutoGPT Platform code |
| Top GitHub repository of 2023 | Yes, by stars added during the year |
For part of April 2023, AutoGPT was adding stars at a faster rate than any project GitHub had previously tracked, a fact that fed back into media coverage and sustained the spike. By the end of 2023, several recap lists named it the top GitHub repository of the year by new stars [2][5][7].
Beyond raw numbers, AutoGPT shifted how mainstream audiences talked about generative AI. Discussion moved from "the chatbot got something wrong" to "the agent ran for two hours, spent $30 in API credits, and produced something half useful." That framing, however imperfect, anchored expectations for the wave of agent products that arrived in 2024 and 2025 [3][4][7].
The ChaosGPT fork from April 2023 captured the cultural moment. A contributor took the AutoGPT codebase, gave it a goal of destroying humanity, and posted Twitter threads of its planning output. The agent did very little real damage, but the demonstration prompted serious public conversation about safety, alignment, and the wisdom of giving an LLM unsupervised access to the open web [4].
The project has faced consistent, well documented limitations from its first days. Many of these limitations apply to LLM agents generally, but AutoGPT was the first system on which they were observed at scale.
| Limitation | Details |
|---|---|
| Hallucinations | The agent invents facts, URLs, file paths, and intermediate results when its context is thin. Without aggressive verification, those errors compound across loop iterations. |
| Infinite loops | Without strong memory, the agent can repeat the same subtask, or oscillate between two near identical plans, until a hard iteration cap fires. |
| API cost | Every loop iteration is a GPT-4 call. A single multi hour run can cost tens of dollars in OpenAI charges, and there have been reports of users running up much larger bills overnight. |
| Context window | Even with GPT-4 Turbo's larger context, AutoGPT can lose track of the original goal as the message history grows. Andrej Karpathy publicly cited the finite context window as the main reason early AutoGPT runs went off the rails. |
| Drift from the goal | Broad objectives lead the model down tangents that look useful in isolation but never circle back to the original task. |
| Code execution risk | Running model generated shell commands or Python scripts on a host machine without a sandbox is dangerous. AutoGPT defaults to Docker for this reason but the option to run locally still exists. |
| Vector store complexity | The original third party vector database integrations were powerful but added a lot of operational overhead, leading the maintainers to drop them later. |
| Reproducibility | Two runs from the same prompt can diverge wildly because of model temperature, tool latency, and changing web content. This makes evaluation hard. |
The 2024 and 2025 releases of the AutoGPT Platform addressed several of these issues, particularly around iteration caps, structured workflows, and benchmark driven testing. The fundamental limits of LLM reasoning, however, remain [3][4][7][9].
In 2024 the project pivoted from a single Python script into a full platform. The new system is structured around two halves: a backend server and a frontend builder. Together they aim to make autonomous agents accessible to non engineers without giving up the openness that defined the original release [8][12].
| Component | Role |
|---|---|
| AutoGPT Server | Core logic, scheduling, agent runtime, marketplace API, and persistence layer |
| AutoGPT Frontend | Visual agent builder, workflow editor, deployment controls, and analytics |
| Forge | Toolkit for writing custom agents with reduced boilerplate |
| agbenchmark | Performance testing harness for measuring agent capability across tasks |
| Marketplace | Catalogue of pre built agents and blocks contributed by the community |
The server is the hard backend: it handles authentication, runs agents on behalf of users, and queues triggers. The frontend is a low code editor where users wire together blocks (search, summarise, extract, send email, post to Slack, and so on) into reusable workflows. Each block is a small unit of capability, similar in spirit to Zapier zaps or n8n nodes, but with LLM reasoning available at every step [8][12].
A central design idea in the modern platform is the continuous agent. Where the original Auto-GPT was a one shot loop you ran from a terminal, the platform supports agents that run on a schedule, on a webhook, or in response to events such as a new file in cloud storage. This shift turns AutoGPT from a sandbox curiosity into something closer to an automation tool that competes with Zapier and n8n, with LLM reasoning baked in [12].
The platform integrates multiple model providers, including OpenAI, Anthropic, Groq, and Llama based local models. This insulates users from any one vendor's pricing changes and lets them route different blocks through different models, for example using Claude for long form writing and GPT-4o for code [12].
The licensing model also changed. The classic agent and most of the repository remain under the MIT License. The newer code in autogpt_platform/ ships under the Polyform Shield License, which permits broad use but bars third parties from offering a competing hosted version of the platform. This dual license arrangement is increasingly common for open core AI infrastructure projects [5][8].
In parallel with the self hosted release, Significant Gravitas runs a cloud beta of the AutoGPT Platform. The hosted version takes care of installation, scaling, and credential management, and is the recommended starting point for users who want to try the platform without standing up their own infrastructure [12].
The AutoGPT name now points at three different things at once, which causes confusion in search results and documentation:
| Name | What it actually refers to |
|---|---|
| AutoGPT (project) | The Significant Gravitas open source autonomous agent and platform |
| AutoGPT Platform | The 2024 onward web application built on top of AutoGPT |
| AutoGPT Agent (Custom GPT) | The third party Custom GPT by agilayer.com listed in the GPT Store |
The Custom GPT covered at the top of this article is the third item. It uses the AutoGPT branding but has no direct connection to Toran Bruce Richards or Significant Gravitas. Anyone looking for the actual open source agent should head to github.com/Significant-Gravitas/AutoGPT or to the project landing page at agpt.co [1][2][5].
The specific system prompt for the agilayer.com AutoGPT Agent Custom GPT is not publicly disclosed in the GPT Store listing. The conversation starters and exposed actions imply a prompt that frames the model as a multi step planner with permission to call browsing, code interpreter, DALL·E, and the Zapier webhook in any combination [1].
No uploaded knowledge files are advertised on the public GPT Store listing for this Custom GPT [1].
hooks.zapier.com
https://agilayer.com/privacy-policy/
{"type":"none"}
{"openapi":"3.1.0","info":{"title":"Zapier Webhook Integration","description":"OpenAPI specification for POSTing to a Zapier webhook","version":"v1.0.0"},"servers":[{"url":"https://hooks.zapier.com"}],"paths":{"/hooks/catch/14137578/3kabw2k/":{"post":{"operationId":"Send Data","summary":"Post Data to Zapier Webhook","description":"This endpoint is used to send data to the specified Zapier webhook.","requestBody":{"description":"Data to be sent to the webhook","required":true,"content":{"application/json":{"schema":{"$ref":"#/components/schemas/WebhookData"}}}},"responses":{"200":{"description":"Successful response"},"default":{"description":"Unexpected error"}}}}},"components":{"schemas":{"WebhookData":{"type":"object","properties":{"email":{"type":"string","format":"email","description":"Email address of the user"}},"required":["email"]}}}}
No public usage guide is published with the agilayer.com Custom GPT listing. For an end to end primer on the broader AutoGPT framework, the Significant Gravitas project documentation at docs.agpt.co is the canonical reference [5][8].
The most useful prompts mirror the conversation starters: marketing strategy generation, logo concepting, research and sentiment analysis, and chart creation from web data. In each case the agent leans on a specific built in tool, with GPT-4 acting as the planner that decides when to switch between them [1].
No public conversation transcripts are bundled with the Custom GPT listing. Demonstrations of the broader AutoGPT framework are widely available on YouTube and on agilayer.com's website [1].