verifiable-orchestration

name	verifiable-orchestration
description	Run multi-agent work as ONE provable causal DAG — parallel subagents, an Orchestrate graph, an immutable spec-seed
version	1
trust	built-in

korgex can split work across subagents and record the whole run as one connected, tamper-evident causal DAG — replayable and auditable (korgex trace/verify). Pick the right shape:

• Independent, flat sub-tasks → the Agent tool. Emit several Agent calls in ONE turn and they fan out concurrently (results return in order; a crashed sibling is isolated). Use for "explore these 3 areas", "review these 4 files".
• Sub-tasks with DEPENDENCIES → the Orchestrate tool. Give it nodes with deps (e.g. explore → plan → implement → review): independent nodes run in parallel, a node waits for its deps, and if a node fails its dependents are skipped rather than run against a broken precondition. The whole run — including the failure topology — is one verifiable DAG.

Lock intent first, for non-trivial runs:

• Pass seed to Orchestrate — the agreed {goal, constraints, acceptance_criteria}. It's recorded as an immutable, hash-chained spec.seed the whole run anchors under, so korgex why/trace walk any result back to the spec it was meant to satisfy and korgex verify proves the spec wasn't altered after the fact. Use it whenever the "what we agreed to build" should be pinned before work starts.

Rules:

• One level deep. A subagent cannot itself spawn agents or orchestrate (hard-enforced) — keep the graph at the top level.
• Each node gets a tool surface scoped to its type (explore/plan/review/research are read-only; code gets the full set).
• Prefer a few well-scoped nodes over many tiny ones; name dependencies explicitly.
• After the run, the result carries root_seq (and seed_seq) — trace from there to audit exactly what each subagent did and why.