Hermes

Self-improving personal assistant. Built-in learning loop creates skills from experience, persistent memory builds a user model across sessions, scheduled automations run unattended, subagents spawn for parallel workstreams, and every run produces a trajectory the eval pipeline can score.

sidecar unreachable

Sidecar state

endpoint: http://localhost:8001

mode: unreachable

reason: sidecar at http://localhost:8001 not reachable

sidecar offline

Start the cycle-97 Hermes sidecar: cd backend/services/hermes_runner && pip install -e ".[agent]" && uvicorn hermes_runner.main:app --port 8001. Override the endpoint by setting HERMES_SIDECAR_URL.

Capabilities

Learning loop

Cluster prior trajectories → propose new skills. Proposed skills land in /admin/observability/agent-os/skills/review and cannot run until approved.

Persistent memory

User model accumulates across sessions. Scoped per user / project / org via lib/agent-os/memory-engine.ts.

Scheduled automations

Cron-style triggers managed by lib/agent-os/scheduler.ts. Every run emits events the audit log captures.

Multi-channel gateway

Telegram / Discord / Slack / WhatsApp / Signal / CLI. Each channel requires explicit operator pairing and an allowlist before activation.

Subagent spawning

Parallel workstreams isolated by sandbox-registry policy. Each subagent inherits the parent run's risk budget.

Trajectory export

Every run produces (input, plan, tool_calls, output, reward_signal) — consumed by the cycle-88+ eval pipeline.

Govern Hermes

Review proposed skills

approve / reject / supersede

Inspect trajectories

JSON or JSONL export

Audit every action

filter by runId / kind / time