Skip to main content
Esta página todavía está en inglés. La traducción al español está en progreso.

​
Heartbeats

Heartbeats — proactive agents with schedule, status, cost Heartbeats are proactive agents. A heartbeat is a configuration that wakes an agent on a schedule so it can triage work without being asked. Instead of you remembering to ask Atlas “any stale PRs?” every morning, Atlas wakes up every 4 hours, checks, and acts only if there’s something to do.

​
The Mental Model

A routine is systematic — the same script runs every day at 7am. A heartbeat is proactive — the agent wakes, looks, and decides whether to act. Every heartbeat run answers a single question:
Given the current state of my domain, should I work now or skip?
The agent returns a JSON verdict. If action: work, the dispatcher lets the agent run its turns. If action: skip, the run ends immediately at near-zero cost.

​
Anatomy of a Heartbeat

Heartbeats live in config/heartbeats.yaml (source of truth) and are mirrored to the dashboard DB. Each entry has:

​
Quick Start

​
1. Create a heartbeat

From Claude Code, use the create-heartbeat skill:
“Create a 4-hour heartbeat for atlas-project that triages stale PRs and issues.”
The skill walks you through the fields, defaults enabled: false, and calls POST /api/heartbeats. The entry appears in the dashboard Heartbeats page as disabled. Alternatively, edit config/heartbeats.yaml directly and call POST /api/heartbeats/reindex to mirror it into the DB.

​
3. Dry-run before enabling

From the dashboard, click Run now on the card (or call POST /api/heartbeats/{id}/run). This fires a one-off manual run regardless of the interval clock. Inspect the result:
  • Did the agent return valid JSON?
  • Was the decision reasonable?
  • Did it stay within max_turns and timeout_seconds?
If yes, flip the enable toggle. The dispatcher will pick up the next interval tick.

​
4. Monitor runs

Every run records started_at, finished_at, status, tokens_input, tokens_output, cost_usd, and the parsed decision_json. The Heartbeats page in the dashboard shows the last 10 runs per card and aggregate 7-day cost.

​
Wake Triggers

A heartbeat wakes on any of the triggers listed in wake_triggers:
  • interval — the scheduler fires every interval_seconds. Almost always included.
  • manual — someone clicks “Run now” or calls the API. Enables operators to debug.
  • mention — a @agent-slug in a ticket comment fires a mention trigger (if the agent has an enabled heartbeat with mention in its wake_triggers). Max 3 mentions per comment.
  • new_task — a goal task created with this agent as assignee_agent wakes the heartbeat.
  • approval_decision — an approval request decided (future feature).
Multiple triggers = more responsiveness. Fewer triggers = more predictable cost.

​
Cost Control

Because heartbeats burn tokens on every run, put guardrails on them:
  1. Start enabled: false. Always. Review the first manual run before enabling.
  2. Use a crisp decision_prompt. If the agent writes a paragraph when it should write JSON, you waste tokens.
  3. Bias toward skip. Train the prompt to skip unless there’s genuine urgency. A 4-hour heartbeat that skips 4 out of 6 runs per day is doing its job.
  4. Cap max_turns low for pure triage heartbeats (5–10). Only raise it for agents that may need to act on multiple items.
  5. Watch cost_7d on the dashboard card. If a heartbeat trends up, the prompt is probably too soft — tighten the skip criteria.

​
Debugging

​
A heartbeat isn’t firing

  1. Check the heartbeat’s enabled — is this one individually on?
  2. Check the dispatcher log (journalctl -u evo-nexus -f or the dashboard service logs) for interval registration.
  3. Call POST /api/heartbeats/{id}/run — does a manual run work? If yes, it’s a scheduler issue; if no, it’s an agent or prompt issue.

​
A run failed

Look at the run’s stderr_tail and stdout_tail. Common causes:
  • Missing required_secrets in .env
  • decision_prompt didn’t force JSON — the parser couldn’t extract the verdict
  • Agent hit max_turns before returning (raise the cap or tighten the prompt)
  • Agent hit timeout_seconds (integration call was slow — raise timeout or add a retry inside the agent)

​
YAML and DB drift

If you edited config/heartbeats.yaml by hand and the dashboard still shows the old state, run:
This rebuilds the DB rows from the YAML. Safe to run anytime.

​
CLI Skills

  • docs/goals.md — link a heartbeat to a goal for context injection
  • docs/tickets.md — @mentions in tickets wake heartbeats
  • Source: dashboard/backend/heartbeat_dispatcher.py, dashboard/backend/routes/heartbeats.py, dashboard/backend/heartbeat_schema.py