Skip to content

GenericGPT.md

Latest commit

 

History

History
172 lines (134 loc) · 8.46 KB

GenericGPT.md

File metadata and controls

172 lines (134 loc) · 8.46 KB
Raw

🤖 GenericGPT Interactive Mode

GenericGPT is a production-hardened autonomous software engineering agent with session persistence, model switching, and multi-provider support.

When you run autogpt with no subcommand or flags, it launches an interactive AI TUI powered by GenericGPT:

autogpt

autogpt-cli.mp4

The interactive shell supports the following commands:

Command Description
<your prompt> Send a task to the GenericGPT autonomous agent
/help Show available commands
/provider Switch AI provider (Gemini, OpenAI, Anthropic, XAI, Cohere, HuggingFace)
/models Browse and switch between provider-native models
/sessions List and resume previous sessions
/status Show current model, provider, and directory
/workspace Show the current workspace path
/clear Clear the terminal
exit / quit Save session and quit

Press ESC at any time to interrupt a running generation.

🔀 Mixture of Providers (MoP)

AutoGPT introduces a high-availability Mixture of Providers architecture. When enabled via the --mixture or -m flag, every prompt is fanned out concurrently to all configured AI providers (Gemini, OpenAI, HuggingFace, etc.). A weighted scoring engine evaluates responses based on:

  1. Length calibration (rewarding detail, penalizing fluff).
  2. Code quality (bonus for language-tagged Markdown blocks).
  3. Structural richness (headings, lists, hygiene).
  4. Reasoning depth (connectivity words and logical flow).
  5. Completeness (punctuation and closing delimiters).

The highest-scored response is selected as the winner and injected into the agent's context, promoting the best "intelligence" available from your configured keys.

The .autogpt Directory

GenericGPT maintains all persistent state inside the workspace root (defaults to the current directory):

.autogpt/
├── sessions/          # Markdown conversation snapshots, auto-saved after every response
│   ├── <uuid>.md
│   └── ...
└── skills/            # TOML lesson files, injected into future prompts automatically
    ├── rust.toml
    ├── web.toml
    └── python.toml

Control the workspace root with AUTOGPT_WORKSPACE:

export AUTOGPT_WORKSPACE=/my/project   # scope all file ops to a specific directory
autogpt

Model Selection

Models are sourced dynamically from each provider's crate. Override the active model without entering the shell:

export GEMINI_MODEL=gemini-2.5-pro-preview-05-06
export OPENAI_MODEL=gpt-4o
export MODEL=<any-model-id>    # global fallback for any provider

How GenericGPT Works

Each prompt travels through a multi-phase pipeline, with every phase reflected live in the terminal UI:

  1. Intent Detection: The agent reads your message and decides whether it can answer directly, call a specific tool, or needs to plan and execute a full multi-step task.
  2. Multi-Provider Fan-out: When mixture mode is enabled, the prompt is sent to all configured AI providers simultaneously and the best response is selected.
  3. Task Synthesis: The agent breaks your goal down into a concrete, numbered list of actionable sub-tasks.
  4. Implementation Plan: A structured plan is generated and displayed, giving you a clear overview of what will be built before execution begins.
  5. Reasoning: Before tackling each sub-task, the agent thinks through its approach, anticipated risks, and the best execution strategy.
  6. Execution: The agent carries out the sub-task by performing file operations, running shell commands, searching the web, calling external tools, and more, atomically and in order.
  7. Build & Verify: If a buildable project is detected, the agent automatically compiles or runs it and attempts to self-correct any failures, retrying up to three times.
  8. Reflection: After completing each sub-task, the agent reviews its own output and decides whether to accept it, retry with corrections, or skip and move on.
  9. Metacognition: The agent tracks outcome patterns across tasks. When it detects repeated failures or inefficiencies, it recalibrates its strategy for the remaining work.
  10. Skill Extraction: At the end of a session, the agent distills domain-specific lessons from what worked and stores them for automatic reuse in future sessions on similar topics.
flowchart TD
    A([autogpt CLI]) --> B{CLI args?}
    B -- none --> C["GenericGPT\nInteractive Shell"]
    B -- "-p / --prompt" --> DP["Direct LLM Prompt\n--mixture for fan-out"]
    B -- subcommand --> SA

    subgraph SA ["Specialized Agent Roster"]
        direction LR
        ARCH[ArchitectGPT] --- BACK[BackendGPT] --- FRONT[FrontendGPT]
        DES[DesignerGPT] --- MGR[ManagerGPT] --- MAIL[MailerGPT]
    end

    C --> PS["Provider & Model Setup\ngemini · openai · xai · anthropic · cohere · hf"]
    PS --> RL["Prompt REPL Loop\nESC → abort_token triggers abort"]

    RL --> IC["classify_intent\nINTENT_DETECTION_PROMPT"]
    IC -- DirectAnswer --> SR["generate_safe\nstream reply to TUI"]
    IC -- ToolCall --> TE["MCP / built-in tool\nMcpCall action"]
    IC -- TaskPlan --> MOP
    SR --> RL
    TE --> RL

    subgraph FTP ["Full Task Pipeline · GenericAgent  ·  AgentGPT base class"]
        direction TB

        MOP{"Mixture of\nProviders?"}
        MOP -- Yes --> MF["Fan-out across providers\nmerge best response"]
        MOP -- No --> SP["Single provider\ngenerate_tracked with token stats"]
        MF & SP --> WS["scan_workspace + walk_glob\nfile-tree snapshot injected as LLM context"]
        WS --> SL["SkillStore.load_for_domain\n.autogpt/skills/domain.toml injected"]
        SL --> SY["Phase: Synthesizing\nnumbered sub-task list"]
        SY --> PG["Phase: Planning\nmarkdown implementation plan"]
        PG --> AP{"Phase: AwaitingApproval\n--yolo to skip gate"}
        AP -- Abort --> ID([Phase: Idle])
        AP -- Approved --> EX

        subgraph EL ["Execution Loop: one iteration per sub-task"]
            direction TB

            EX["Phase: Executing n / N"]
            EX --> RS["Reasoning\nReasoningResult: thought · approach · risks"]
            RS --> AR["LLM emits ActionRequest JSON array"]

            AR --> AD{Action type}
            AD -- "CreateFile / WriteFile\nPatchFile / AppendFile" --> FW[Filesystem writes]
            AD -- "ReadFile / ListDir\nFindInFile / GlobFiles" --> FR["Filesystem reads\nwalk_glob + pattern_matches"]
            AD -- RunCommand --> SH["Shell execution\ncwd + timeout"]
            AD -- GitCommit --> GC[git stage and commit]
            AD -- WebSearch --> WEB[DuckDuckGo search]
            AD -- McpCall --> MCP["MCP server tool\nstdio / SSE transport"]
            AD -- MultiPatch --> MPT[Atomic multi-patch]

            FW & FR & SH & GC & WEB & MCP & MPT --> RC["ActionResult\nstdout · stderr · success"]

            RC --> BV{"Build artifact\ndetected?"}
            BV -- "Cargo.toml / package.json / Makefile" --> TB["auto-build\nretry up to 3 · self-fix prompt on failure"]
            TB -- pass --> RF
            TB -- fail --> RS
            BV -- No --> RF

            RF["Phase: Reflecting\nReflectionResult: outcome · corrective_actions"]
            RF -- Retry --> RS
            RF -- "Skip / Success" --> MC

            MC{"mta feature\nenabled?"}
            MC -- No --> NT
            MC -- Yes --> MTE["Phase: MetaCognizing\nrecord_task_outcome → MetacognitionEntry\nshould_adjust_strategy?"]
            MTE -- ok --> NT
            MTE -- adjust --> MH["METACOGNITION_PROMPT\nstrategy hint injected for next task"]
            MH --> NT
        end

        NT{"More sub-tasks?"}
        NT -- Yes --> EX
        NT -- No --> FU{"Follow-up tasks\nneeded?"}
        FU -- Yes --> SY
        FU -- No --> SE

        SE["Skill Extraction\nextract_lessons → .autogpt/skills/domain.toml"]
        SE --> SS["Session walkthrough\n.autogpt/sessions/"]
    end

    SS --> RL
Loading