agent-evals-and-observability

Agent Evals and Observability

Evaluation asks whether behavior meets a defined criterion on a declared dataset or production sample. Observability supplies traces, logs, metrics, correlations, and diagnostic context. Use both; neither proves what the other does.

Multi-agent evaluation boundary

Use multi-agent-coordination to define the team's ownership, communication, intent and integration contracts; return here for comparable datasets, graders, trajectory evidence and uncertainty. Evaluate both individual and combined behavior. Add cases for constraint loss at synthesis, peer approval outside scope, stale handoffs and shared-practice recurrence.

Measure coordination activity alongside accepted integrated outcomes, review cost and task mix. Rising messages or approvals with flat delivery is a diagnostic signal, not proof that checks should be removed. Necessary high-assurance review is a near miss, not a detector failure. A verified answer cannot compensate for an authority violation. Do not claim a universal ceremony threshold.

Workflow

1. Define the decision, risk, task contract, trajectory contract, and unacceptable outcomes. Select evidence by harm, reversibility, and deployment stage, not a staged completeness scale.
2. Create an immutable dataset version and manifest before comparing versions. Declare provenance, rights/consent, slices, fixtures, expected side effects, contamination risk, limitations, retention, and changelog.
3. Select complementary graders that observe the claimed property. Use deterministic checks for observable mechanics; use execution/environment checks for state and side effects; use human, model-judge, pairwise/ranking, domain, and safety review where appropriate.
4. Run the candidate and baseline under comparable conditions. Preserve run configuration, stochastic repeats where variability affects the decision, failures/timeouts, and trajectory evidence.
5. Report a multidimensional profile and uncertainty. Use paired comparisons where possible; inspect slices, missingness, base rates, confounders, effect sizes, and multiple comparisons. Non-significance is not equivalence.
6. Apply a risk-tiered release gate: hard safety, privacy, authorization, and side-effect invariants cannot be averaged away. Record authority, insufficient-evidence outcomes, rollback, and follow-up. Load release-engineering when this evidence must be incorporated into artifact promotion, deployment, rollback, or a broader release train.
7. Instrument production with minimized, redacted telemetry. Feed verified incidents and near misses into cases after consent, transformation, and contamination review.

When an evaluation needs intervention detail, keep the existing aggregate count separate from optional typed events. Absence of event detail means it was not collected; it does not mean zero. Collect only fields needed for the question, exclude raw messages and hidden reasoning by default, and do not treat lower counts alone as success. See trajectory review.

Stop when the supported decision, evidence gaps, residual risks, and responsible owner are recorded. Escalate rather than infer a pass when required evidence is unavailable or conflicts.

Load By Need

Need	Load
Idle-agent capacity claims, event interleavings, callback recovery or hibernation traces	event-runtime-evaluation
Compare task-relevant discovery, external artifacts, delegated consumption, or fresh-session recurrence	references/agent-interface-enrichment.md
Validate a lesson candidate before promotion and check recurrence	references/lesson-validation.md
Choose evaluation evidence and contracts	references/evaluation-design.md
Separate performance, diagnostic, and replay evidence; connect offline to online measures	references/evidence-modes-and-metric-mirrors.md
Build immutable cases, fixtures, and provenance	references/datasets.md
Select or calibrate graders	references/graders.md
Evaluate monitor independence, correlated failures, detection, or escalation delivery	references/monitor-quality.md and templates/monitor-assessment.md
Validate judge stability, abstention, and escalation	references/judge-stability.md
Check evidence cherry-picking, overreach, or bounded in-workflow probes	references/grounding-probes.md
Provide requested machine-readable evidence or join multiple evaluation artifacts	references/decision-audit-record.md and templates/decision-audit-record.json
Evaluate grounded claims and attribute RAG failures	references/grounding-attribution.md
Define measures or compare runs	references/metrics-and-statistics.md
Evaluate a bounded Choice/Score/Noul decision model or GLiNER classifier inside an agent	system-one for its question contract, held-out labels, calibration where probabilities exist, and model-level latency; use this skill for the agent's end-to-end task and trajectory evidence
Review tools, state, recovery, or side effects	references/trajectory-review.md
Record optional, privacy-bounded human intervention detail	references/trajectory-review.md#optional-intervention-events
Triage a regression or decide release readiness	references/regression-and-release.md
Design traces, logs, metrics, or privacy controls	references/production-observability.md
Interoperate with OpenTelemetry	references/opentelemetry-genai.md
Need framework-specific implementation	references/framework-routing.md
Exercise the methodology with safe probes	references/synthetic-scenarios.md
Benchmark an agent's Supabase task competence with the official harness	supabase agent evals reference
Verify a source claim or status	references/source-index.md

Templates

Use templates/eval-plan.md, templates/interface-enrichment-comparison.md, templates/dataset-manifest.md, templates/grader-specification.md, templates/trace-review.md, templates/run-report.md, templates/release-gate.md, and templates/metric-mirror.md. They are decision records, not checklists that manufacture evidence.

Use grounding probes only when the decision depends on evidence-backed claims. Check material qualifications and whether evidence supports the proposed action, not just the quoted sentence. Preserve separate findings, probe errors, and unknown outcomes; a probe score cannot authorize action.

Monitor and oversight evaluation

Evaluate the supervisor as well as the agent. Load the monitor-quality reference when the decision depends on detecting conduct violations, monitor independence, or reliable escalation. Keep agent behavior, monitor detection, and delivered escalation as three separately evidenced outcomes. A task-correct result can still violate its trajectory contract; a raised alert can still fail to reach its owner. Self-assessment is diagnostic testimony, never the sole oracle for the agent's conduct. Different models alone do not establish independent oversight.

Protected-control evidence

For attempts to alter supervision, load control-integrity evaluation. The controlled agent is neither sole approver nor sole evidence source. Evaluator datasets, labels, oracles, and gate configuration are protected assets; editing them to manufacture a pass invalidates the result. Missing, degraded, redirected, or suppressed mandatory controls are control-integrity incidents. Assess task success, conduct, control state, and escalation separately, using independent effect and configuration evidence. Route runtime hold/containment to agent-production-operations and rights or exceptions to ai-governance.

Guardrails

• Do not use a scalar score, keyword hit, schema validation, or repeated output as proof beyond the property it directly observes.
• Do not capture prompts, outputs, tool arguments, intermediate reasoning, credentials, or personal data by default. Minimize before export, redact early, restrict access, set retention/deletion paths, and prepare incident response.
• Do not treat human judgment as definitive truth. Human and model graders require explicit rubrics, calibration, disagreement analysis, and revalidation when conditions change.
• Keep prevalence-oriented production samples separate from risk-enriched challenge cases. Do not silently reweight either into the other.

When Not To Use

Use systematic-debugging to investigate one active defect before proposing fixes. Use verification-methodology for general evidence-backed completion claims, secure-software-engineering to design preventive controls, and framework skills for SDK-specific code. To benchmark an agent's competence on Supabase tasks with the official scored scenarios, use the supabase agent evals reference — this skill covers methodology, not a named harness.