AgentLab

AgentLab is a Pi-inspired agent harness playground for exploring how to build one in Go.

The project will experiment with Go-based agent orchestration, a Bubble Tea terminal UI, and sandboxed tool execution.

Hello World

AgentLab currently has an Ollama hello-world command backed by a small in-memory session abstraction:

go run ./cmd/agentlab

Pass a custom prompt with --prompt or as positional text:

go run ./cmd/agentlab --prompt "Search the sandbox for aurora, then tell me the reported status."

By default, AgentLab snapshots testdata/smoke-sandbox and exposes these read-only Ollama tools to the model:

• list_files
• read_file
• search_text

Run the local tool smoke test against the configured model with:

make smoke

Run the Pi-based smoke test in an isolated Docker container with:

make smoke-pi

The Pi target copies only testdata/smoke-sandbox into a temporary container workspace, enables Pi's read-only read, grep, and ls tools, and writes Pi's Ollama config to a temporary directory. It does not commit your Ollama endpoint/model. Set PI_OLLAMA_BASE_URL and PI_OLLAMA_MODEL to override them; otherwise the script reads the configured Ollama provider settings.endpoint and settings.model from AGENTLAB_CONFIG or ~/.config/agentlab/config.yaml. Set PI_OLLAMA_PREFLIGHT_TIMEOUT to override the default 15s model-list timeout, PI_OLLAMA_PREFLIGHT_CHAT_TIMEOUT to override the default 45s warmup chat timeout, and PI_SMOKE_TIMEOUT to override the default 180s Pi prompt timeout. Set PI_OLLAMA_PREFLIGHT_CHAT_TIMEOUT=0 to skip the warmup chat, or PI_SMOKE_SKIP_PREFLIGHT=1 to skip all Ollama preflight checks. Set PI_SMOKE_EVENT_LOG=0 to disable the default live Pi event log.

The CLI is built with Cobra and accepts --config for an explicit YAML config file:

go run ./cmd/agentlab --config ./config.yaml

Configuration is loaded from ~/.config/agentlab/config.yaml on macOS by default. On other platforms it uses the standard user config directory. Set AGENTLAB_CONFIG to point at a different config file.

default_provider: local

providers:
  - name: local
    type: ollama
    settings:
      endpoint: http://localhost:11434
      model: qwen3-coder
      context_window: 98304
      # Optional. Use true/false for most thinking models, or low/medium/high for models that support levels.
      think: true

  - name: openai
    type: openai
    # Optional if OPENAI_API_KEY is set.
    api_key: sk-...
    settings:
      model: gpt-5.4
      # Optional. Defaults to https://api.openai.com/v1.
      base_url: https://api.openai.com/v1

Use a non-default configured provider with:

go run ./cmd/agentlab --provider openai --prompt "Say hello from AgentLab in one short sentence."