Recursive Agent Harness (RAH) - MAIster

RAH is MAIster’s governed pattern for recursive multi-agent work. An orchestrator can delegate bounded child Runs, wait without occupying an active agent session, collect typed results, and reduce them into one result that the normal evidence and review path can evaluate.

Reference delivery shape

orchestrator → writer → independent judge → human review → promotion
      └──── read-only agent or Flow researchers ────┘

The reference shape deliberately has one worktree writer. Research children are read-only and publish schema-validated public Run results. The coordinator collects those results by Run identity rather than scraping assistant prose.

Governance boundaries

RAH preserves the properties that hidden subagents usually lose:

every child is a first-class Run in a visible recursive tree;
depth, fan-out, active-child count, total child count, token, time, and failure budgets are bounded and snapshotted;
agent and Flow children use the same admission queue;
result schemas, revisions, producer identity, validity, and artifact manifests are durable;
failures route through bounded rework or human escalation;
independent verification and the standard readiness gate still apply before promotion.

Result-only research

A result-bearing research Flow that changes no code can finish as Done after publishing a valid result. It does not create a meaningless diff-review step. A Run that writes code still follows the normal Review and promotion path.