Harness Engineering: The Operating System for Agentic AI
B · Useful, limited Short BCG practitioner article with a usable five-part control model for agents, but its performance figures come from unnamed BCG client projects with no baseline or method disclosed.
Summary based on the full text·Original: about 8 minutes to read
In short
Aparna Kapoor, Vincent Paca and Nerijus Zemgulys of BCG Singapore (16 September 2026) argue that agents fail in large companies because of missing controls around them, and that the model itself matters less. They call the fix "harness engineering": an operating layer with five parts that sets scope, rules, monitoring, context and quality checks. They claim it produced large gains at a Southeast Asian bank and a software company, and advise firms to start small, build their own harness and pick workflows carefully.
Key numbers
- 30%+uplift in wealth-advisor revenue productivity at a Southeast Asian bank (BCG client claim)
- 3xtime advisors spent actively engaging clients on the agentic platform
- 5xengineering efficiency gain claimed for the bank build and for a software client's shift to agent-native delivery
- 50%faster design-to-deployment on the bank platform
As reported by the source.
English·Web page·opens bcg.com
What it reports
- Most large companies are not actively scaling agents, and those that are report limited bottom-line effect (citing unnamed recent surveys).
- At a large Southeast Asian bank, an agentic advisory platform let advisors triple client-facing time, with a 30%+ uplift in wealth-advisor revenue productivity and a four-to-five-times improvement in customer conversion.
- The same build reported five times higher engineering efficiency and 50% faster design-to-deployment; the authors attribute this to the harness (evaluation, approvals, guardrails) more than the model.
- A software company moving from coding assistants to a full harness across the product life cycle raised team productivity "by as much as five times".
- Five components: specs (machine-testable scope contract), constitution (deterministic rules on what agents must and must never do, escalation triggers, named accountable person), control panel (audit trail, escalation matrix, recovery), context hub (the private data agents need), and quality gates.
- Four kinds of quality gate: automated checks, critic agents evaluating worker agents, human stage gates with a named approver, and regulatory compliance gates.
- A "maturity ladder" moves content from full human approval (for example client-facing investment proposals) to spot checks as trust builds. In banking the authors observed the junior base of the workforce pyramid shrinking.
How they got it
Consulting experience from BCG engagements, described as case examples. No sample, baseline, measurement period or independent verification is given.
Limits
- All performance numbers are BCG's own client claims, unnamed and unaudited.
- "Build, don't buy" aligns with BCG's interest in selling build services.
- No failure cases or costs of building the harness are reported.
Why it matters for companies
The five-part checklist is a practical way for a non-technical manager to ask whether an agent project is ready: is there a written scope, hard rules, an audit trail, the right internal data, and named people at the approval points. The point that quality gates differ by risk, and that approval levels should loosen only as evidence builds, applies to any process where errors are costly. The observation that junior roles thin out under this model is relevant to workforce planning. Treat the multipliers as illustrations from a seller, and ask for baselines before using them in a business case.
Summary based on the full text. Read the original for the full argument and data.Read the original on bcg.com
- Tags
- AI agents