◆ DARWIN · self-learning engine
← Genius cockpit
Autonomous · Propose-only · Kill-switched

DARWIN — autonomous self-learning engine

Senses signals from the world, proposes improvements under hard safety limits, and learns from every run. A human approves before anything ships.

synchronizing… · connecting
System
—
awaiting state
⏻ Kill switch — when off, the monitor and brain no-op. No sensing, no proposing.
—
awaiting state
▶ Execution arm — when on (and an API key is set), approved proposals run via the CodeBuild runner. Default off.
Monthly budget
$—
spend tracked per run · resets monthly
—
Signals sensed
—
Proposals pending review
—
Runs executed

The learning loop

A closed cycle. Phase 1 runs SENSE → PLAN; the dashed stages light up in Phase 2 once the runner is enabled.

01
SENSE
Ingest signals — releases, docs, community, metrics.
→
02
PLAN
Turn signals into scored, classified proposals.
→
03
ACT
Approved safe changes get applied by the runner.
Phase 2
→
04
EVALUATE
Measure result, cost, and expected vs actual value.
Phase 2
→
05
LEARN
Feed outcomes back to sharpen the next cycle.
Phase 2

Live signals

Most recent first — what DARWIN is sensing right now.

Signal feed

—
📡Waiting for the first signals…

Pending proposals

Human-in-the-loop. Approve to queue for Phase 2, or reject to dismiss.

🧩No proposals yet. DARWIN proposes as it learns.

Recent runs

Execution history with cost and outcome. Populates once the Phase 2 runner is live.

⚙️No runs yet — the Claude Code runner activates in Phase 2.

Memory · what DARWIN has learned

Semantic playbook (distilled nightly by the curator) + recent episodic lessons written back by runs.

📘 Playbook
—
🧩 Recent lessons
No lessons yet — runs write these back via the DARWIN MCP server.