Agents and Orchestration

Every agent task opens a HarnessRun. The run binds to a work item, thread, session, and optional workspace. Agent runs produce typed receipts, artifacts, and approval records that are cross-device shared truth.

Runs can be paused, resumed, and replayed. Child runs created through delegation carry parent_run_id for full provenance tracking.

Agents is the surface profile for autonomous agent orchestration. The CEO Nome coordinator frames missions, assigns roles, and manages delegation chains.

Squads are groups of agents working toward a shared objective. Squad routing determines which agents participate and how results aggregate.

Every risky action can be approval-gated. Approval records include the reason, proposed action, tool tier, and expiration. They are cross-device shared truth, not surface-local memory.

Delegation creates child runs with explicit parent_run_id. A run never silently forks into a second identity.

Every external action emits a typed ToolReceipt with tool name, tier (read/write/danger), redacted inputs, output summary, duration, and success status.

Receipts attach to work items via canonical refs. The tool policy matrix defines per-tool governance: max calls per run, approval thresholds, offline behavior, cost weight, and allowed roles.

Hermes implements a continuous self-improving loop. When a task sequence is novel or valuable, it generates a reusable SKILL.md procedure.

Skill retrieval uses progressive disclosure — only a compressed skill index is loaded initially. Full skill content is injected only when the reasoning engine determines it's needed for the current objective.

Gastown is the topology and runtime reference layer for agent coordination. Canonical refs are wired through the capability spine and surfaced in the capability manifest.

Gastown governs how agents are distributed across execution topology without replacing the product-facing coordinator identity.