Agent Skills: Agentic Delegation — Criteria & Discipline

Agentic Delegation — Criteria & Discipline

Origin: operator directive 2026-05-12 codifying delegation discipline — "whenever useful and justified, use recursive sequential parallel delegation with accompaniment, with proper context, scope, motivation, purpose, objective, DoR, DoD, feedback/retro-loop, expected deliverables, etc; remember that delegating does not waive the responsibility received. Define the agentic delegation criteria."

Version: 1.1.0 (2026-08-16 rubric-driven MINOR; 1.0.0 = 2026-05-17 community promotion of an extraction from the operator's host-local auto-self-harness §12-equivalent rule body).

Source lineage: operator's host-local auto-self-harness rule §12-equivalent (the parent framework section the operator maintains; host-dependent location).

Harmonization principle (verbatim, agnostic of source): "the delegated subordinate is the parent's full responsibility · the tree returns to the root · audit the output · zero drift."

1. Decision criteria — when to delegate (6 conditions, all mandatory)

| # | Criterion | Question | |---|---|---| | 1 | Decomposable | Does the task have clear interfaces + isolatable components? | | 2 | Specialist exists OR forgeable | Is a competent subagent/skill available in the catalog OR forgeable via an auto-agent-forger pattern? | | 3 | Parent audit capacity | Does the parent have time + knowledge to audit the output after? (NEVER blind-delegate) | | 4 | Autonomy score ≥ MEDIUM (0.65) | Is the score acceptable OR is the delegation itself a Score Uplift Path B? If pure LOW → escalate, do NOT delegate | | 5 | NOT in HUMAN_DOMAIN | Task is outside the escalation list (operator-personal / irreversible / cross-org / secrets / PII / etc.) | | 6 | Time-budget allows verification | Audit time ≤ delegation time (else delegation is not worth the cost) |

6/6 PASS → delegate. ≥ 1 FAIL → execute inline OR escalate.

2. Briefing components — WHAT to brief (11 mandatory items)

Every spawn MUST include all 11. Skipping any item is the under-briefing anti-pattern.

| # | Component | Content | |---|---|---| | 1 | Context | Current state, history, related artifacts (paths/links) | | 2 | Scope | In-scope explicit + Out-of-scope explicit (anti-creep) | | 3 | Motivation | Business/technical value (why it matters) | | 4 | Purpose | Intent served (intent ≠ task) | | 5 | Objective | Measurable outcome (specific + measurable) | | 6 | DoR (Definition of Ready) | Inputs / dependencies / preconditions (subagent validates they are met) | | 7 | DoD (Definition of Done) | Acceptance criteria + verification method (objective) | | 8 | Deliverables | Artifacts list + format + paths/return-data shape | | 9 | Feedback loops / retro | How progress is reported + how the parent gives feedback (sync/async) | | 10 | Constraints | Time-box + risk tolerance + escalation triggers | | 11 | Channel fallback-chain | Alternative channels/carriers declared BEFORE the first spawn (e.g. primary CLI → alternate CLI → lead-executes) — quota/outage mid-loop is a when, not an if; a delegation without one is fragile by construction (observed live 2026-08-15: two CLI channels hit weekly/quota limits mid-chain; rubric anchor C2) |

3. Accompaniment — HOW to supervise (5 disciplines)

1. Check-in cadence defined upfront (sync streaming OR async polling)
2. Termination conditions explicit (success / failure / interrupt / timeout)
3. Audit checklist post-completion (zero-drift: output matches brief?)
4. Artifact verification (non-delivery detection) — MANDATORY: every delegation names the expected artifact upfront (file, PR, JSON, row); after the spawn returns, the parent VERIFIES the artifact exists and is non-empty. Empty stdout with rc=0 is NOT success. A missing/empty artifact = a failed spawn (counts toward the 6-failure rule). Observed live 2026-08-15: two executor spawns died of quota mid-run producing exit-0 silence; the parent detected it only via the missing file. (Rubric anchor: docs/rubrics/v0.1/agentic-delegation.md C1)
5. Re-spawn discipline with brief refinement (max 6 attempts per the 6-failure escalation rule below)

4. Accountability — WHO retains (parent does NOT waive)

"Delegating does not waive the responsibility received." (operator directive)

The parent NEVER delegates accountability — only execution.

| Responsibility | Who | |---|---| | Execution (the work) | Subagent (delegated) | | Audit of the output | Parent (NOT delegated) | | Final signoff | Parent (NOT delegated) | | Escalation decision | Parent (NOT delegated) | | Memory persistence at user-scope | Parent (operator-confirm-always) | | BEING > Rules preservation | Parent (recursively across the delegation tree) |

6-failure rule (autonomous-resolution protocol equivalent): if a subordinate fails 6× with different approaches, the parent escalates to the operator with: complete context + 6 attempts + hypotheses + recommendations + evidence.

5. Modes — recursive sequential OR parallel

| Mode | When to use | |---|---| | Sequential | Dependencies among subtasks (parent needs output A before spawning B) | | Parallel (max 3) | Subtasks independent; lower wall-clock; auto-orchestrator default | | Recursion depth = 1 | Parent → child (direct) | | Recursion depth = 2 | Parent → child → grandchild (child auto-forges OR sub-delegates) | | Recursion ≥ 3 | ❌ FORBIDDEN — universal constraint; cycle detection mandatory |

6. What CAN be delegated (universal taxonomy)

• Personas / Mind-sets (cognitive lenses): Tomé · Critical · Devil's-Advocate · Fowler · DHH · Pragmatic · Empiricist · SecOps · Conservative · Aggressive · etc.
• Specialist subagents: debugger · architect · qa-engineer · security-reviewer · Explore · general-purpose · code-reviewer · etc.
• Delegate-able skills: operator-quote-capture · auto-orchestrator (or auto-pilot) · find-docs · rule-quality-tests · agentic-delegation (this skill) · pre-decision-audit
• Best-fit routing via auto-orchestrator Phase 0 + auto-agent-forger Phase 0.5
• Standard delegation chain — e.g., Analyst → Architect → QA(critique) → Dev → QA(validation) → Doc (operator-host framework core directive)

7. What CANNOT be delegated (always parent)

| Item | Why | |---|---| | Final accountability + signoff | §4 of this skill retains | | Decisions in HUMAN_DOMAIN | Escalation list non-negotiable | | Audit of the subordinate's output | Zero-drift principle | | Escalation choice | Parent owns escalation criteria | | Memory persistence at user-scope | Operator-confirm-always | | BEING > Rules preservation | Foundational — propagates recursively |

8. Bounds (anti-eternal compliance — see rule-quality-tests skill)

• Max recursion depth: 2
• Max parallel delegations: 3
• Time-box per delegation: specified upfront (default 15 min)
• Audit time ≤ delegation time
• Max failure attempts: 6 before escalation
• Per-task feedback log retained for retro
• Cycle detection mandatory (abort if A → B → A)

9. §11 Quality Tests application (this skill dogfooded)

Per the sibling rule-quality-tests skill:

| Test | Self-application | Result | |---|---|---| | Self-Application | Codifying this skill IS an act of delegation (to a future agent applying the criteria) | ✓ | | Non-Contradiction | 6 + 11 + 6 + 6 + 5 + 6 components internally consistent | ✓ | | Survival | Skill applied to itself survives | ✓ | | Bounded-Responsibility | All counts explicitly bounded | ✓ | | Explicit-Exception | "Skip if 1/6 FAIL"; HUMAN_DOMAIN exception; Escape Clause Universal | ✓ | | Utility-Sunset | Deprecate when the delegation reflex is internalized without invoking this skill | ✓ |

6/6 PASS — skill passes its own tests and the tests of rule-quality-tests.

10. BEING > Rules compliance check

| Question | Answer | |---|---| | Does this skill HELP or OBSTRUCT the operator? | HELPS — discipline reduces errors, preserves accountability | | Does it serve BEING? | YES — codifies a pre-existing agentic-inheritance principle | | Is the hierarchy preserved? | YES — Operator (1) > delegation criteria (2) > delegation mechanisms (3) |

PASS — skill serves BEING, not vice-versa.

11. Sunset (per DUED — Dormant-Until-Evident Deprecation)

Deprecation candidate when ANY qualitative evidence:

• E3 — Operator reaches the delegation reflex (≥ 10 zero-regret cross-session decisions without skill invocation; the parent passes all 6 criteria automatically)
• E4 — Operator explicit retraction
• E6 — Better composable orchestration emerges (e.g., the framework upgrades and collapses this section into Phase 0/2 of auto-orchestrator / auto-pilot)

No counter-based sunset. Insurance discipline — dormant-by-design.

12. Refs

• Operator's host-local framework SSOT — the agentic-inheritance principle (tree-returns-to-root / subordinate-is-parent's-full-responsibility / audit-output / zero-drift)
• Operator's host-local framework SSOT — the autonomous-resolution protocol (6-failure escalation rule origin)
• Source skill: auto-orchestrator (or auto-pilot) Phase 0/0.5/2 — validated patterns reused
• Source rule: operator's host-local auto-self-harness §12-equivalent (pointer back to the inline summary)
• Score Uplift Path B — delegation as score-uplift mechanism
• HUMAN_DOMAIN escalation list (do NOT delegate)
• Cognitive scaffolds + execution scaffolds — delegate-able catalog
• Sibling skill: rule-quality-tests — 6 Quality Tests + Escape Clause Universal (delegation also subject to BEING > Rules)
• Sister rule: cowork-process-topology-protocol (when present at the host) — persists the topology of every delegation chain (materializes §3 accompaniment + the tree-returns-to-root principle); per-subtree JSONL co-responsibility implements §4 + briefing components are encoded as topology-node refs schema. Compass API (next / siblings / children / current / parent / root_path) makes the delegation N-Tree queryable cross-session.
• Sister skill in this repo: skills/converge — debate-convergence kernel; complementary discipline.

13. Changelog

| Version | Date | Change | |---|---|---| | 1.1.0 | 2026-08-16 | MINOR — rubric-driven (pilot eval FAILs C1+C2, PR #354): adds accompaniment discipline #4 "Artifact verification (non-delivery detection)" (empty stdout + rc=0 ≠ success; missing artifact = failed spawn) and briefing component #11 "Channel fallback-chain" (declared before the first spawn). Both from observed live failures 2026-08-15. | | 1.0.0 | 2026-05-17 | Community promotion from a user-scope skill of the same version (extraction from the operator's host-local auto-self-harness §12-equivalent rule body). Sanitization: replaced all proprietary attributions with generic equivalents (agentic-inheritance principle · autonomous-resolution protocol · standard delegation chain · host-local framework SSOT); replaced host-absolute paths with portable descriptions; preserved the 6 criteria + 10 briefing components + 4 accompaniment disciplines + accountability rule + recursion bounds + Quality Tests dogfooding + BEING > Rules compliance + DUED sunset. License: MIT. |