references/agentic_delivery_governance.md
Agentic Delivery Governance
The accountability layer behind delivery_loop_gate.py. When agents execute delivery
work, the failure mode is not bad output — it is unowned output: work no human is
accountable for, verified by nobody, closed by the optimism of the thing that did it.
The 2025–2026 vendors converged on the same governance shape; this file encodes it.
The delegation model (why G1/G2 exist)
- Linear's shipped design: issues can be delegated to agents, but the human stays primary assignee; the agent is a contributor. Delegation transfers execution, never accountability. → Gate G1: every task names a human owner.
- Atlassian Rovo (GA 2026): agents are assignable and @mentionable inside Jira, but
"every action remains logged and auditable", and multi-step plans pause for human
oversight at decision points. → Gate G2: agent-executed tasks name a human reviewer;
hard rule: no un-reviewed
transitionJiraIssueto Done, no permission changes inside a loop.
Verification discipline (why G3/G4 exist)
- Machine-checkable definition of done: the reliable loop shape is plan → act →
verify with a deterministic verifier → reflect. A criterion without a command or a
threshold ("looks good", "improved") cannot gate anything. → G3: acceptance = a
cmd, or a criterion containing a measurable number. - Never trust self-report: the agent that did the work does not adjudicate the work.
Evidence precedes status:
donewithout recorded evidence is rejected (G4) — the same invariant as agent-harness's "a verify pass without --evidence is exit 6" and autoresearch-agent's locked evaluator ("never modify the gate you are judged by").
Terminal-state honesty (why G5/G6 exist)
From the loop-library contract: loops end in a named terminal state — success, clean no-op, blocked, approval-required, exhausted, stagnated. Two corollaries the gate enforces:
- Close is refused while any task is neither done nor waived (G5); waivers are human decisions with recorded rationale.
- An exhausted budget (attempts or iterations) is an escalation, never a success report (G6). Budgets are first-class: max attempts per task, max loop iterations, and the stop fires mechanically, not when the agent feels finished.
Design principles for PM loops
- Workflows before agents (Anthropic): most PM automation is a routed workflow (classify → run tool → report). Reach for the autonomous loop only when the task needs fresh feedback each cycle — flow snapshots, retro follow-through, RAID hygiene.
- Evaluator-optimizer needs clear criteria: the generator/critic loop pays off exactly when acceptance is machine-checkable — which is why the gate forces G3 before any loop starts.
- Context from structured interfaces, not prompt-stuffing: the agent's Jira context comes through the MCP (Teamwork-Graph-style structured access), snapshotted to a file the loop can re-read — every iteration is executable by a fresh session.
- Agent-readiness is a data-hygiene property: agents amplify the Jira they are given (DORA 2025's amplifier finding). Field completeness, acceptance criteria in descriptions, and honest statuses are prerequisites, not nice-to-haves.
Sources
- Linear, "Agents in Linear" / Linear for Agents — https://linear.app/agents and https://linear.app/docs/agents-in-linear
- Atlassian Rovo — https://www.atlassian.com/software/rovo and Rovo agents docs https://support.atlassian.com/rovo/docs/agents/
- Anthropic, "Building Effective Agents" (orchestrator-workers, evaluator-optimizer, simplicity-first) — https://www.anthropic.com/research/building-effective-agents
- Forward Future, Loop Library (terminal-state taxonomy; "never report an error or
exhausted budget as success") — vendored at
loop-library/SKILL.md - DORA, State of DevOps 2025 (AI as amplifier) — https://dora.dev/dora-report-2025/
- This repo:
engineering/agent-harnessreferences/verification_discipline.md (reward-hacking failure mode; locked evaluators) - SiliconANGLE, "Atlassian opens Teamwork Graph, pushes Rovo agentic execution" (Team '26 coverage) — https://siliconangle.com/2026/05/06/