[Work Governance / Research] Calibrate adaptive scouting, decomposition, and delegation policy

[Sequela02]

Objective

Determine and benchmark when Kiln should keep scouting and execution in the primary agent, delegate a bounded investigation, or use a coordinated multi-agent topology. Replace the current assumption that non-trivial work benefits from orchestration by default with a measured, provider-neutral decision policy if the evidence supports that change.

This issue authorizes research, fixtures, and evaluation design. It does not authorize changing the production default before the promotion gate passes.

Related work

• [Skills / Cross-harness] Add portable orchestration and research workflows backed by Kiln governance #75 delivered the portable orchestration-workflow; it did not establish a general quality, latency, token, or cost advantage.
• [Architecture/Research] Add native bounded-work authority for scope fidelity and overengineering control #19 owns bounded-work authority and semantic overengineering measurement.
• [Architecture / Long-term] Make Kiln configuration schema-first and onboarding-guided #72 owns configuration schema, descriptors, effective state, and guided mutation; it does not own the coordination algorithm.
• docs/architecture/core/work-governance.md and docs/architecture/coordination/coordination.md describe the current canonical policy and runtime topology.
• docs/research/active/adaptive-work-governance.md owns the evidence synthesis and hypotheses for this issue.

Decision questions

1. Should the primary agent perform a minimal repository orientation before any delegation?
2. Which measurable task properties justify a delegated scout: search breadth, context pressure, independent questions, specialization, latency, or another signal?
3. When does delegated scouting with the same model or a cheaper model improve success, total cost, latency, or retained root context?
4. What is the correct decomposition unit: independently testable and reviewable result, file count, estimated duration, or another contract?
5. Which dependency, write-overlap, tool-density, and shared-state conditions require direct or sequential execution?
6. When does a reviewer add value beyond deterministic tests, typecheck, lint, builds, invariants, and diff inspection?
7. Which context envelope should a child receive: full fork, curated contract, or isolated task plus durable resources?
8. Which fan-out, depth, retry, and budget limits prevent coordination pathologies without suppressing useful parallelism?

Current evidence

The evidence is conditional rather than supportive of one universal posture:

• OpenAI recommends starting with a single agent and adding multi-agent complexity only when needed: https://openai.com/business/guides-and-resources/a-practical-guide-to-building-ai-agents/
• Anthropic recommends Explore-Plan-Code and suggests subagents for particular investigations, while its multi-agent research results apply chiefly to breadth-first research and used substantially more compute: https://www.anthropic.com/engineering/claude-code-best-practices and https://www.anthropic.com/engineering/multi-agent-research-system
• Google Research evaluated 180 configurations and found gains on parallelizable tasks but 39?70% degradation on a sequential planning benchmark: https://research.google/blog/towards-a-science-of-scaling-agent-systems-when-and-why-agent-systems-work/
• TeamBench reports conditional team value and a 49% false-approval rate among verifier approvals that failed its deterministic grader: https://arxiv.org/abs/2605.07073
• ContextBench finds only marginal context-retrieval gains from sophisticated scaffolds over 1,136 issues from 66 repositories: https://arxiv.org/abs/2602.05892
• Practitioner evidence disagrees in useful ways: Simon Willison documents context-preserving Explore subagents, while Armin Ronacher reports poor results outside investigative tasks. These are operational observations, not controlled comparisons: https://feeds.simonwillison.net/guides/agentic-engineering-patterns/subagents/ and https://lucumr.pocoo.org/2025/7/30/things-that-didnt-work/

No searched source provides a controlled repository benchmark comparing primary-agent scouting with same-model and cheaper delegated scouts. Kiln must measure that locally.

Required benchmark

Use real, replayable Kiln tasks stratified by dependency shape, write overlap, tool count, repository familiarity, expected context volume, and single-agent baseline competence.

Compare:

• scouting: primary-agent orientation, same-model delegated scout, cheaper-model delegated scout;
• execution: direct/sequential, centralized manager-worker, independent parallel work;
• verification: deterministic oracles only, reviewer plus deterministic oracles;
• context: full fork, curated task envelope, isolated task plus resource references.

Run both fixed-budget and unconstrained observations, with at least 3?5 repetitions per admitted cell. Record deterministic task success first, then escaped defects, total provider tokens by class, coordinator turns, latency, retries, rework, merge conflicts, duplicated exploration, context/handoff loss, false approvals, child count/depth, and operator intervention. Unknown usage remains unknown.

Do not run meaningless Cartesian cells. Pre-register exclusions such as parallel mutation of a shared file or a reviewer condition without an external oracle.

Candidate policy shape

A candidate may be evaluated with these provisional rules:

• direct execution is the baseline;
• the primary agent performs only enough orientation to frame bounded questions and retain integration responsibility;
• delegate scouting only for specific independent questions or context-heavy searches;
• parallelize only independent work with non-overlapping mutable ownership;
• use a centralized integrator when child outputs affect one final decision;
• treat reviewer output as a hypothesis; completion requires available external oracles;
• decompose by independently completable, testable, and reviewable outcome rather than a fixed file or LOC threshold;
• cap fan-out, depth, retries, and total coordination budget.

These are hypotheses, not accepted defaults.

Acceptance criteria

• Repository fixtures and strata are published with exact candidate/config/model/harness identities.
• Baseline, candidate, and relevant ablations are replayable.
• The scout comparison includes primary, same-model child, and cheaper-model child conditions.
• Quality, cost, latency, coordination overhead, rework, and failed/omitted runs are reported.
• Deterministic oracles remain separate from LLM reviewer judgments.
• The decision policy uses measurable task properties and reports uncertainty or unsupported capability.
• Any default change includes migration/deletion of the superseded path, without a compatibility variant.
• Work Governance architecture, coordination architecture, instruction doctrine, skills, config descriptors, and native projections are updated together only after promotion.
• Independent benchmark-readiness and architecture review have no unresolved high or medium findings.

Non-goals

• No universal requirement to delegate scouting.
• No universal preference for a cheap scout.
• No fixed task size expressed solely as files, LOC, or minutes.
• No mandatory Planner-Executor-Verifier topology.
• No inference that more agents, more tokens, or more reviews imply better work.
• No transfer of coordination-policy ownership into Roadmap 12 or skill prose.

[Sequela02]

Implemented the architectural/default-policy slice on dev in 83d2775.

What changed:

• Direct execution is now the baseline.
• Removed file-count/risk heuristics and the generic multi-file orchestration trigger.
• Orchestration now requires an explicit configured trigger plus a task-structure source of expected value (independent parallel work, specialization, context isolation, independent evidence, or authority/security separation).
• LLM review remains hypothesis generation; deterministic/external verification owns completion evidence where available.
• Consolidated duplicated workflow evidence lists into the canonical evidence matrix.
• Updated the adaptive-work-governance research and architecture docs.

This commit does not claim the benchmark/promotion gate in this issue is complete. The policy change is an explicit operator-directed Sequel doctrine decision, not a measured product-efficiency claim. This issue should remain open for replayable cross-model/harness calibration, cost/latency/quality measurements, and ablations of direct vs delegated topologies.