SoluteLabs’ Agentic Harness: verification debt is the real bottleneck
Oct 12 SoluteLabs field note: INFORM → SPEC → BUILD → VERIFY after a three-line Tailwind change broke sticky nav — named failure modes include verification debt when AI produces faster than the team can prove safe.
Source: The Agentic Harness: How We Rebuilt Our SDLC for AI-Augmented Product Engineering
SoluteLabs rebuilt their SDLC after a three-line Tailwind change broke position: sticky across nav and header — and the fix lived in one engineer’s head. The Agentic Harness is the operating system they put around AI-generated work: INFORM → SPEC → BUILD → VERIFY, each phase with a defined artifact and a human checkpoint. The failure modes they name are the ones always-on factories hit first: context hallucination (agents “clean up” protective weirdness), self-confirming code (same agent writes and validates), silent regressions, and verification debt — AI produces faster than the team can prove the output is safe to ship.
INFORM turns briefs and incidents into requirement IDs with a traceability matrix and checks CLAUDE.md in as operational memory — constraints learned from real failures, including the sticky-nav cleanup pattern. BUILD runs on Claude Code; for non-trivial diffs an independent agent review runs before a human sees the PR. VERIFY is a dedicated review pass separate from the implementer, plus visual regression, production probes and revert-detection canaries, and human sign-off on evidence before release. Architecture, security-sensitive work, and irreversible decisions stay with named human ownership.
Why it matters for code review
Verification debt is the review bottleneck dressed as a throughput win: more diffs arrive than evidence can keep up with. SoluteLabs’ answer is phase gates and a second-pass reviewer before humans, so continuous agent output lands as inspected work rather than first-draft volume. Pair with Reed’s Evidence Ledger and draft-PR-or-BLOCKED exits and Firmo’s second-model signed review.