Loops
A loop is a scheduled agent behavior that runs without human intervention. Loops are what make the mesh autonomous: while you sleep, they're watching, consolidating, proposing, and maintaining.
Two tiers of loops run on every mesh:
- Supervisor loops — run on supervisor nodes; concerned with mesh-wide health and cross-node coordination.
- Memory loops — run on supervisor nodes; concerned with the shared mesh memory.
On top of those, you can define custom loops for your team.
Supervisor loops (selected examples)
These are representative — the full list is visible in the mesh dashboard.
| Loop | What it does |
|---|---|
| Node Health Monitor | Watches for stale nodes and restarts failing services |
| Cost Tracker | Aggregates per-node spend for the dashboard |
| Cost Forecaster | Projects monthly spend and flags anomalies early |
| Idle Session Reaper | Stops agent sessions that have been idle past a threshold |
| Daily Report | Summarizes the mesh's activity and proposes follow-ups |
| Proposal Triage | Groups and prioritizes inbound proposals |
| Lineage Drift Detector | Watches for divergence between apps and the appgrammars that built them |
| Trust Score Audit | Reviews external dependencies and ratings |
Each loop runs on its own schedule. Scheduling details and on/off toggles are configurable per-mesh.
Memory loops
Memory loops keep the mesh's collective knowledge healthy.
| Loop | What it does |
|---|---|
| Memory Consolidation | Merges near-duplicates and decays stale entries |
| Memory Distribution | Propagates high-value facts across the mesh |
| Daily Proposal Synthesis | Turns observations into actionable proposals for human review |
Custom loops
Define your own loops for team-specific behaviors — nightly security scans, weekly dependency upgrades, monthly cost reviews. A custom loop is an agent task plus a schedule plus a target (a specific node, all workers, or the supervisor).
Create a custom loop from the dashboard: Loops → New Loop. You pick the schedule, the target, and the prompt. The loop runs on your LLM keys like any other agent work.
Interaction with proposals
Most loops don't change your system directly — they produce proposals for human approval. See Tasks, Reports & Proposals. This keeps autonomy and safety in balance: the mesh can notice everything and act on nothing without your say-so.
Enabling loops
During mesh setup, you pick an initial set of loops to enable. You can enable or disable each loop individually from the dashboard at any time. Disabling a loop stops its schedule; it does not cancel any proposals or reports it already produced.
Next
- Tasks, Reports & Proposals — Where loop output lands for review
- Observe — See loops run in real time