Turn a goal into finished work, on its own.
Agentryx is built without the usual manager-and-workers metaphor — no personas, no fixed roles, just tasks, skills, and evidence. A goal either runs directly or splits into a sub-task DAG, the same rule at every level, and a failed step gets replanned with the actual error in hand — never a blind retry. Nothing is marked done because an agent says so; a goal contract tracks what would prove it, and completion is audited against evidence.
A task either runs directly or splits into a sub-task DAG, by the same recursive rule at every level. Single-agent execution isn't a special path; it's the base case of that rule.
When a sub-task fails, its parent replans with the error and partial results in hand. If it still can't recover, the failure propagates up and a higher level replans with broader context.
Every goal gets an implicit contract: what would prove it done, and what evidence currently covers that. A worker's self-report isn't enough — completion is audited against test results, checks, and artifacts, and for benchmark-matched goals, an independent adversarial run has to pass against the real thing before it can be marked complete.
Run-or-split, run, synthesize — implemented and in daily use. A survey of roughly 19 agent frameworks found the combination of uniform recursion, per-node synthesis, and upward-propagating replan-with-understanding uncommon; the closest validated peer is Anthropic's orchestrator-workers pattern, single-level where this is recursive.
Every root goal gets a durable contract — objective, acceptance criteria, an evidence ledger. A worker can propose the work is done; only an audit against tests, checks, and artifacts can mark the goal complete. Weak or missing evidence leaves it open, not "probably fine."
One run finished "complete" with checks passing while the actual result wasn't usable — the completion signal was self-certified. Now, for benchmark-matched goals, an independent adversarial run has to pass against the real artifact before completion, and a non-trivial task can't reach done without an independent review.
“An agent can say the work is done. Only the evidence gets to agree.”