agentic-tend Multi-Agent Evidence
Use evidence-separated agent contexts through observable port and feedback boundaries. The public multi-agent evidence model owns the effective-theory rationale; this skill realizes its conditional port contract.
Establish the evidence boundary
Activate this skill only when another context can:
- form a claim from a distinct validity source or method without inheriting another conclusion;
- keep applicable evidence held out from a proposed implementation;
- seek a discriminating failure without receiving the success narrative it should challenge.
When inputs, methods, evidence, and failure hypotheses cannot be usefully separated, continue with one context.
Specify the motivation, inputs, outputs, withheld information, provenance obligation, and feedback edges that change observable task behavior.
Show the plan before delegation
Once the skill activates, the coordinating agent presents a human-visible plan before delegating. Expose the motivation and observable outcome, the ports that current evidence requires, their input and output boundaries, intentionally withheld information, feedback and stopping evidence, authority, and unvalidated claims.
After the plan is visible, continue within the current contract and authority. Pause for a user-owned semantic choice, expanded authority, destructive action, or another boundary that requires human input.
Establish the minimum observable port topology before filling local execution detail: motivation, contract, port boundaries, and applicable validity sources constrain each port's realization. Structure, execution, and feedback may interleave after the boundary is coherent.
Derive conditional ports
Derive the minimum effective port graph from current task data.
- Coordinator bridge: consumes the request, established contract, task data, authority, and returned evidence. It exposes the plan and delivery, bounds port interfaces, maintains feedback edges, and applies the stopping contract. Transported claims retain their proper validity sources, and disagreement follows the cheapest discriminating source.
- Prior-evidence port: appears only when valid evidence can be formed independently and implementation exposure could contaminate it. Give it motivation, contract, current facts, and applicable validity sources, but withhold the proposed implementation when independence matters. It returns supported mechanical predicates, judgment-changing findings, and explicit unvalidated boundaries.
- Worker port: consumes a bounded objective, contract, authoritative facts, allowed actions, and applicable capabilities. It returns an artifact or claim, observations, evidence, provenance, and unresolved boundaries. Its internal trajectory remains unspecified.
- Posterior-review port: appears when an independent attempt can expose a plausible failure or compress review. Give it the resulting artifact, contract, evidence, and needed environment without an unnecessary success narrative. It returns a counterexample, disagreement, evidence gap, or bounded support with provenance.
The coordinator is the stable human ingress and egress once multi-agent execution is active. Prior-evidence and posterior-review ports appear with their stated motivation. Add, combine, repeat, or retire ports as observations update task data.
Preserve oracle and capability ownership
A mechanical predicate belongs in prior evidence when an available oracle can decide it and the contract makes the check useful. An interpretation traces a result to motivation, contract, evidence, or human authority and remains a claim. A claim with no deciding validity source retains an unvalidated status.
When software work activates $software-engineering, that capability owns implementation topology and realization, testing, oracle integrity, and delivery semantics. This skill decides whether those judgments benefit from separate contexts while preserving every selected capability's semantic owner.
Route recursive feedback
Transport observations with the input or condition, expected contract behavior, observed result, validity source, and provenance. Pass interpretations only when the next reader needs them to judge a live alternative.
Posterior review may recurse like a limited cache hierarchy: a cheaper local layer absorbs routine detail, while a failure, disagreement, or audit gap escalates. Add a layer when its independent evidence can change judgment, and forward judgment-changing distinctions while preserving pointers to accessible artifacts.
Agreement establishes consistency within the compared inputs and methods. Correctness remains attached to proper validity sources, including shared models, specifications, sources, and tools that can create common-mode failures. Route disagreement to the cheapest discriminating source or probe. Return user-owned semantics to the user, and retain unresolved claims as unvalidated.
Stop when the established contract has sufficient evidence from its proper validity sources and no material unresolved conflict remains within the authorized scope.
Deliver the effective result
Present motivation and observable outcome before agreement, disagreement, evidence, and unvalidated claims. Keep design, implementation, checks, falsification attempts, and provenance reachable through available files, tool outputs, and lane evidence.
An unavailable underlying transcript or artifact produces an explicit audit gap. Follow the active authority boundary for any external action.