BCG’s harness OS: four unequal quality gates, not one green check
Sep 16 Kapoor/Paca/Zemgulys: specs, constitution, control panel, context hub, and four gate types (automated / critic / human stage / regulatory) — maturity ladder for when humans still must say yes.
Source: Harness Engineering: The Operating System for Agentic AI
BCG’s Sep 16 piece (Kapoor, Paca, Zemgulys) argues the reliability crisis around agents is an operating-system problem: models commoditize; the harness captures how this org works. They map five components to a PC OS metaphor — specs (machine-testable scope contract), constitution (deterministic always/never/escalate rules with named accountability), control panel (audit trail, escalation matrix, failure recovery), context hub (private shared memory so agents stop guessing), and quality gates.
The review-relevant cut is that quality checks are not equal. Four gate types:
- Automated — programmatic failure prevention.
- Evaluation — specialized critic agents score worker agents.
- Human stage — a named person marks an activity complete.
- Regulatory — a quality agent enforces compliance standards.
A maturity ladder then moves work down the clearance policy as trust grows: client-facing proposals may require every output reviewed; battle-tested internal drafts can become spot-check. Humans relocate to supervisors who set goals and boundaries up front and intervene at critical checkpoints; specialized critics handle more routine gates. Receipts they cite: a Southeast Asian bank harness that lifted wealth-advisor productivity 30%+ and sped design-to-deployment ~50%; a software company shifting from coding assistants to harness engineering across the product lifecycle claiming up to 5× team productivity.
Their adoption rules are anti-hype: start small, build don’t buy the org-specific harness (context and quality signals are the moat), and be selective — only workflows with documented work and checkable quality. Productive complication for lights-out factories: BCG still puts human stage gates and regulatory gates on the critical path; continuous momentum is not “skip review,” it is routing the right gate type to the risk.
Why it matters for code review
Under agent PR volume, one green CI check is a category error. Continuous harnessed throughput needs the four-gate split — deterministic blocks, independent critics, named human stage gates, and compliance — plus a maturity ladder so review bandwidth tracks blast radius. The FOMO that agents aren’t working overnight is wasted if the morning queue has no constitution, no critic, and no named acceptor.