agentcore

Go-native agent runtime for managed-agent sandboxes.

agentcore is an experimental, embeddable runtime for running coding agents in headless sandboxes. It focuses on a small native binary, explicit lifecycle semantics, typed events and low per-session overhead.

Current scope

• Multi-turn sessions with model/tool loops.
• Typed bidirectional session events.
• Bash and filesystem-oriented coding tools.
• Context compaction with tool-call pairing preservation.
• Cancellation, tool timeouts and panic containment.
• Isolated asynchronous sub-agents (result delivery is still evolving).

MCP, production WebSearch, durable task storage and a stable public API are not implemented yet. See the capability status before relying on compatibility claims.

Quick start

git clone https://github.com/yanpgwang/agentcore
cd agentcore
go build -o agentcore ./cmd/agentcore

./agentcore \
  --session-id local-1 \
  --agent-config config.json

Example config.json:

{
  "system_prompt": "You are a helpful coding agent.",
  "model": "claude-sonnet-4-6",
  "initial_prompt": "List all TODO entries in the project.",
  "permission": "always_allow"
}

Events are written to stdout as newline-delimited JSON. Set ANTHROPIC_API_KEY before running. ANTHROPIC_BASE_URL can override the official Anthropic endpoint for a compatible gateway.

Documentation

The full documentation is hosted at yanpgwang.github.io/agentcore.

• Getting started
• Runtime architecture
• Lifecycle semantics
• Event reference
• Development and testing

Documentation changes belong in the same commit as the behavior they describe. Public Go APIs are documented in source and rendered by pkg.go.dev.

Development

go test ./...
go test -race ./...
go vet ./...

The project is intentionally small. New packages and abstractions are added only after a concrete boundary has been demonstrated by more than one implementation.

License

MIT