README.md

metadata

title: agent-loop — FORGE Self-Improvement Orchestrator
emoji: 🔄
colorFrom: blue
colorTo: red
sdk: docker
pinned: true
license: mit
short_description: Feedback loop — trace→learn→propose→approve→deploy

🔄 agent-loop

FORGE Self-Improvement Orchestrator

Closes the feedback loop. Runs on a configurable cron interval, detects underperforming agents, generates improved system prompts via NEXUS, creates drafts in agent-prompts, and notifies you for approval. Never auto-deploys — every change requires human approval.

Improvement pipeline

TRACE ──(events)──▶ LEARN ──(scored)──▶ LOOP
                                          │
                              NEXUS ◀──(generate proposal)
                                │
                          PROMPTS ◀──(POST draft, status=draft)
                                │
                        RELAY ◀──(Telegram notify)
                                │
                       YOU ──(approve via dashboard or API)
                                │
                    PROMPTS persona updated → agents fetch at next startup

Decision thresholds

Condition	Action
avg_reward < REWARD_THRESHOLD (0.2)	Generate improvement proposal
error_rate > ERROR_ESCALATE (15%)	Mark critical, escalate via RELAY
reward_delta < 0.05 after 24h	Mark cycle inconclusive
reward_delta > 0.05 after 24h	Mark cycle successful

REST API

POST /api/cycle                     Trigger a cycle manually
GET  /api/cycles                    List recent cycles
GET  /api/proposals                 List proposals (state, agent filters)
POST /api/proposals/{id}/approve    Approve → deploy to agent-prompts
POST /api/proposals/{id}/reject     Reject proposal
GET  /api/health/agents             Agent health snapshot
GET  /api/stats                     Loop statistics

MCP

GET /mcp/sse · POST /mcp

Tools: loop_status, loop_trigger, loop_proposals, loop_approve,
       loop_reject, loop_health, loop_cycles

Secrets

Key	Description
LOOP_KEY	Optional write auth (X-Loop-Key header)
LEARN_URL	https://chris4k-agent-learn.hf.space
TRACE_URL	https://chris4k-agent-trace.hf.space
PROMPTS_URL	https://chris4k-agent-prompts.hf.space
NEXUS_URL	https://chris4k-agent-nexus.hf.space
RELAY_URL	https://chris4k-agent-relay.hf.space
CYCLE_MINUTES	Default: 60
REWARD_THRESHOLD	Default: 0.2
CYCLE_ENABLED	Set false to pause

Built by Chris4K — ki-fusion-labs.de