Iteration
Iteration in matic is not a fallback mechanism. It is a first-class design pattern embedded at every level of the runtime, from within-run validation retries to org-wide hypothesis-driven evolution. The system closes feedback loops across temporal and organizational scopes so refinement remains durable, traceable, and human-supervised. This section covers the protocols, cycles, and governance structures that turn continuous refinement into an operational discipline rather than an ad hoc response to failure.
Probe-Decision Cycle
How the runtime evaluates agent-owned probes — lightweight, side-effect-free heuristics describing capability, confidence, bandwidth, and availability — to make staffing and routing decisions without invoking full reasoning capacity.
Feedback Loop
The interlocking feedback mechanisms that connect outcomes to goals, experience to skill development, performance to probe updates, and post-delivery observability to continuous monitoring.
Revision Cycles
How work items, planning documents, and probes move through structured revision — from validation failure paths and grounding check retries to human-driven document markup and provisional probe refinement.
Amendment Propagation
How confirmed hypotheses and human decisions flow through the system as PRs — proposing new work items, policies, skills, routines, or goal changes — with every consequential change recorded as an immutable decision at the appropriate scope.
Exit Criteria
The conditions that must be satisfied before work completes at each level: work item acceptance criteria and floor coverage, engagement-level learning and assessment gates, team releasing protocols, and project phase transitions.
Blocking Review
How HITL checkpoints pause execution for human judgment — including checkpoint definition points, escalation paths when recipients are unavailable, the dual-write rule for decision records and audit entries, and team-level blocking conditions.
Retry and Recovery
Run-level failure detection and state preservation, automatic reassignment when matching agents are available, daemon-level recovery on restart including signal replay and dirty worktree handling, and scheduled retry policies for routine-based work.