WardenAI is an agent governance layer for software delivery — structured human checkpoints, domain-specific skills, and a self-improving correction loop that stops the same mistake from repeating across sprints.
Four things a framework alone doesn't hand you: a decision record, a memory for mistakes, portable domain knowledge, and a cost curve that doesn't punish rigor.
Every phase transition needs a human APPROVE / REJECT / DECIDE, recorded to EVIDENCE.md.
Two recurring corrections in a sprint become a proposed SKILL.md rule before the next ticket runs.
Composable SKILL.md files carry the rules — not tribal knowledge stuck in one person's head.
Mechanical roles run on cheaper, faster models; judgment-heavy roles run on the stronger one — not a blanket top-tier default.
Multi-agent frameworks solve agent-to-agent handoffs. They don't solve what happens around those handoffs — quality gates, domain knowledge, human oversight, and a way to stop the same mistake from happening twice.
Agents don't pass work directly to each other. Every phase produces output, emits a structured checkpoint, and waits for a human APPROVE / REJECT / DECIDE before the next phase runs. The human acts at gates — never mid-phase.
Seven agent roles. Composable domain skills. A BDD traceability chain from requirement to test. And a self-healing loop that tries to fix its own failures before surfacing them.
On a red test, the DEV agent reads the stack trace, makes a targeted fix, and re-runs — up to 3 iterations. If still failing, Gate 4 surfaces a diagnosed failure with iteration history, not a raw traceback. The loop has hard stops: secrets, scope expansion, or repeated identical fixes bypass it entirely.
Each agent loads composable SKILL.md files: an Azure Function skill enforces
_require_env() for secrets, Service Bus error isolation, PII rules, and golden example tests.
Skills are separate from role definitions — the same domain skill can compose across multiple agents.
PO writes acceptance criteria as Given/When/Then scenarios. TL verifies them for technical completeness. DEV implements step definitions against the same file. CR checks that every AC has a corresponding passing test. A traceability gap at any step is a Gate block.
After Gate 3, the context-compactor skill distils the accumulated ticket into a DEV-BRIEF.md
— just the file structure, edge cases, fixtures, and reference values DEV needs. The full ticket stays as
the audit record. DEV reads the brief. Context noise drops, reliability goes up.
Every agent emits a standard checkpoint before handoff: ALWAYS SURFACE items, conditional FLAG items,
and AUTO-BLOCK conditions with no bypass. The human responds with APPROVE,
REJECT: <reason>, or DECIDE: <choice>. Every response is recorded
in EVIDENCE.md — the committed human oversight audit trail.
Agent frontmatter declares a model: tier. ARCH, TL, CR, and QA run on Sonnet
(judgment-heavy, cross-cutting decisions). PO and Post-Process run on Haiku (structured formatting,
categorisation). Any orchestration layer picks this up automatically, so spend scales with how much
judgment a step actually needs.
Post-Process reads self-heal logs, CR findings, and QA results. When two corrections of the same type appear in a sprint, the system proposes a concrete rule addition to the relevant skill file. The human approves the exact text. It gets committed. The loop closes.
=== CORRECTIONS === section. Appends one index row to root
CORRECTIONS.md — ID, ticket, sprint, category, agent, status.=== CORRECTIONS === section from
the 2–3 referenced tickets. Drafts a concrete addition to the relevant SKILL.md — an exact line, not
vague guidance.Always guard optional API response fields with a null check before property access.
CORRECTIONS.md (root) ├─ CORR-001 — full detail ├─ CORR-002 — full detail ... 80 more entries ├─ Pattern Analysis table └─ Proposal blocks Skill-refinement reads the entire file every run. Merge conflicts on every ticket.
CORRECTIONS.md (root) │ 6 columns, append-only rows │ no prose, no detail tickets/TICKET-007.md │ === CORRECTIONS === │ full detail, self-heal ref └─ no other ticket can conflict Pattern detection reads index. Proposal fetches 2–3 tickets. Scales to any sprint size.
Not every team needs eight gates. This is built for the ones where "the tests pass" isn't the same thing as "this is safe to ship."
Where a reviewer needs to see why a change is correct, not just that CI is green — and where that reasoning has to be committed somewhere auditable.
One skill system and one gate protocol reused across squads, instead of every team hand-rolling its own agent prompts and quality bar.
If a correction is worth flagging twice, it's worth writing into the skill file once. The loop is built to make that automatic instead of aspirational.
WardenAI is a private, proprietary system. Get in touch for a technical walkthrough of the governance layer, skill definitions, and correction pipeline.