Execution Harness
Use this skill when an agent workflow needs an auditable envelope around work performed by deterministic tools, local models, cloud models, or other agents. It is a contract and verification layer, not an executor and not an authority escalation mechanism.
Choose it when
- a task can degrade because of hardware, budget, context, or availability;
- several agents should explore independently before sharing findings;
- candidate artifacts need a protected last-known-good state;
- a handoff must state what completed, what remains, what changed, and what is still uncertain;
- model performance is being compared across different harnesses or hardware;
- commercial workflows must distinguish reusable code/architecture from model weights carrying non-commercial restrictions.
Do not use it as a substitute for domain-specific validation. A passing harness record only means the declared invariants were respected; it does not prove the underlying scientific, security, artistic, or product claim.
Required execution record
Capture, when applicable:
- mission and immutable-enough context snapshot;
- capabilities actually available and constraints in force;
- requested execution state and actual execution state;
- every material delta between those states, with reason and acknowledgement;
- evidence references and whether they were independently verified;
- artifacts and their digests;
- consequences on touched resources, including reversibility;
- uncertainties and unresolved evidence;
- last-known-good recovery point;
- remaining work and a bounded handoff.
No silent degradation
If requested_state != executed_state, the workflow must record an
ExecutionDelta with a non-empty reason and explicit acknowledgement. Running a
shorter, cheaper, lower-quality, different-model, reduced-context, or otherwise
altered job and presenting it as the requested job is invalid.
A workflow may still choose a degraded path when policy permits it; it must say what changed. This rule is intentionally inspired by mature workflow skills that make resource-driven fallbacks explicit rather than silently changing the task.
Last-known-good rule
Treat new output as a candidate. Promote it to the recovery point only after verification evidence exists. A failed candidate remains useful experimental evidence but must not replace the last-known-good artifact.
Divergent exploration
For high-uncertainty work, prefer:
isolated exploration -> evidence gate -> cross-pollination -> independent verification
Agents should not receive each other's hypotheses during the isolated phase unless the task explicitly requires collaboration. Only candidates with evidence and a passed evidence gate enter the cross-pollination packet. A producer cannot serve as its own independent verifier.
This mode complements, rather than replaces, the existing strategic outsider, red-team, and domain-specific verification paths.
Capability Atlas
Record observations by at least:
task x model x harness x hardware x quantization
and attach evidence references. Latency, quality, outcome, and failure mode are observations, not universal properties of the model. Never collapse two results that used materially different harnesses into a model-only benchmark.
The initial implementation is deliberately in-memory and serialization-friendly. Persistence should reuse the repository's established bounded/local data stores rather than introduce a second memory authority.
Commercial licence boundary
Keep code/architecture reuse distinct from weight usage. Non-commercial model weights must not be promoted into a commercial execution path merely because the surrounding implementation is permissively licensed. Record the actual weights licence in the execution context and fail closed when commercial use conflicts with it.
validate_commercial_weights_policy checks a caller-supplied declaration. For
commercial use, it recognizes NonCommercial and BY-NC across case, whitespace,
hyphen and underscore variants. It rejects empty declarations and placeholders
unknown, unspecified, unqualified, tbd, n/a, none and NOASSERTION.
Treat a return without an exception only as absence of those declared blockers,
never as legal clearance. Qualify the actual weight terms against their sources
before any future use, including custom terms and licence expressions: this
helper performs no licence discovery or permission verification. With
commercial_use=False, it makes no rights decision either.
Asymmetric budget guidance
Context ingestion, retrieval, deliberation, and final generation need not use the same resource class. Prefer cheap deterministic/context-compaction stages first, then spend larger-model or cloud budget only when verification, uncertainty, or output quality justifies it. Record the selected path and any divergence from the requested execution state.
Current boundary
Version 1 exposes contracts only. It does not automatically route, execute, persist Atlas observations, activate models, modify memory permissions, promote artifacts, or merge code. Integration with existing router, local harness, trajectory, memory, and handoff surfaces should happen incrementally behind their existing policy gates.
