AutoGPT
Type: full-code · Vendor: Significant Gravitas · Language: Python, TypeScript · License: MIT (classic/) + Polyform Shield (autogpt_platform/) · Status: active · Status in practice: mature · First released: 2023-03-16
Provide a platform to create, deploy, and run continuous autonomous AI agents that chain LLM reasoning with a fixed Command/Tool catalogue, an episodic action history, file/web/code-exec components, and an optional human-approval gate.
Description. AutoGPT is the founding 'autonomous-agent' project — the prototype that turned ReAct + a fixed command palette + a goal into a viral category in March 2023. The current top-level repo ships two products: 'AutoGPT Platform' (the new low-code block builder in autogpt_platform/) and 'AutoGPT Classic' (classic/original_autogpt/), the original ReAct loop. The classic loop composes Agent + Forge components: ActionHistory (EpisodicActionHistory), Watchdog, CodeExecutor (Docker-sandboxed), WebPlaywright, WebSearch, FileManager, Todo, Skill (SKILL.md), and UserInteraction.
Agent loop shape. ReAct-style cycle: each step the prompt strategy assembles system prompt + episodic action history + watchdog hints + context, sends to LLM, parses ActionProposal, optionally pauses for user approval before executing the chosen Command. EpisodicActionHistory records every proposal-result pair for replay/resume. WatchdogComponent guards against runaway loops; --continuous-limit caps total iterations. TodoComponent maintains an explicit task list. The new platform reframes this as a directed graph of blocks edited in the Agent Builder UI.
Primary use cases
- continuous autonomous agents over a fixed Command catalogue
- low-code block-graph agent builder (autogpt_platform)
- Docker-sandboxed code execution + web automation
- todo-driven goal pursuit with episodic history
Key concepts
- Agent + Forge components → react (docs) — Agent(BaseAgent) composed of pluggable components (Action History, Watchdog, Code Executor, Web, File Manager, Todo, Skill, User Interaction).
- EpisodicActionHistory → agent-resumption — Every (proposal, result) tuple persists for resume/replay.
- Continuous mode + continuous_limit → step-budget — Opt-in autopilot with iteration cap; watchdog watches for runaway loops.
- CodeExecutorComponent → code-execution — Per-agent Docker sandbox (agent_id_sandbox).
- TodoComponent → todo-list-driven-agent — Explicit task-list driver.
- AutoGPT Platform blocks → event-driven-agent (docs) — Graph builder: each block performs a single action.
Patterns this full-code implements —
- ★★Code Execution
CodeExecutorComponent with per-agent Docker sandbox.
- ★★Event-Driven Agent
AutoGPT Platform expresses agents as directed graphs of blocks, each performing a single action; blocks can be triggered by external sources.
- ★★ReAct
OneShotAgentPromptStrategy is the canonical ReAct strategy.
- ★★Step Budget
--continuous-limit caps iterations; WatchdogComponent watches for runaway loops.
- ★Todo-List-Driven Autonomous Agent
TodoComponent is a first-class component on the Agent.
- ★★Tool Use
Command system; function_specs_from_commands exports OpenAI function specs.
- ★★Agent Resumption
EpisodicActionHistory persists proposal-result tuples; resume requires reloading the history.
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Instantiates
Alternatives & relatives
Listed as alternative by (2)
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified