The Build Lens

BuildAugust 25, 20263 min readby Sav Banerjee

Build an Agent Harness: 4 Inputs That Get AI Agents Into Production

Only 11% of organizations have an AI agent running in production. The gap is not engineering — it is four inputs nobody wrote down before the build started. Here is the checklist, and what each input becomes inside the system.

Agent HarnessForward Deployed StrategistForward Deployed EngineerEval HarnessAgentic AIBerkeley RDIAgentic AI Summit 2026RAGMCPAI GovernanceRegulated IndustriesCX StrategyData StrategyMeasurementEnso LabsStrategy to Ship

TL;DR

11%

of organizations have an agent in production

inputs decide whether yours is one of them

49.6%

recall@1 after the harness changed — same model

171%

average ROI on the agents that do cross

The blocker is not the model. Leaders name evaluation gaps, governance friction and reliability — all three are harness problems, all three trace to a missing input.

Four strategy artefacts are the input spec. The brief, the journey map, the measurement plan, the segmentation. Each maps to exactly one harness component.

Write them before the build, not after. Reverse the order and you join the 89% that never ship.

The Campanile at UC Berkeley at dusk with San Francisco and the Bay beyond — Part 2 of The Forward Deployed Strategist, on building an agent harness

An agent harness is everything in an agent that is not the model — prompt construction, tools, memory, permission boundaries, validators, and the loop that decides whether any of it worked. Part 1 argued the scarce half of forward deployment is the strategist seat. This is what that seat actually hands over.

Why agents stall at the pilot

The crossing, animated — 38% of organizations piloting AI agents against only 11% with one running in production

Deloitte's Tech Trends 2026 puts 38% of organizations piloting and 11% in production. Gartner still expects 40%+ of agentic projects cancelled by 2027. Every named blocker is something a strategist decides, not something an engineer builds.

The model is the cheapest part of the system. The expensive part is deciding what it is allowed to do, and what would count as it working.

The 4 inputs that decide it

The input/output spec, animated — the brief becomes the system prompt and permission boundary, the journey map becomes the tool allowlist and escalation gate, the measurement plan becomes the eval harness and golden-set CI, and the segmentation becomes the retrieval strategy and recall target

None of these are background documents. Journey mapping was always a measurement discipline wearing a design costume — you were documenting where value leaked and which decision at which moment would have changed it. Those are exactly the decisions an agent has to encode.

In a regulated build each one also becomes a control: audit trails on the runtime, NIST AI RMF or FDA and MLR governance over the tool list and corpus, human approval before anything irreversible. The escalation rule stops being a design preference and becomes the thing an auditor asks to see.

The proof: the harness moved the number

Same model, different harness — Microsoft's AgenticRAG results, animated: BRIGHT recall@1 rising from a 27.8% embedding baseline to 49.6%

Microsoft's AgenticRAG paper held the model constant and changed only how the system was allowed to look for evidence — a retrieval-strategy decision, made upstream, by whoever defined what "relevant" means for the business. It also hit 92% answer correctness on FinanceBench, within two points of being handed the right evidence outright.

For a Fortune 500 manufacturer we encoded a lead scientist's relevance rules before a single pipeline stage was written. For a pharma AI Center of Excellence, the claims and governance layer came first and the agents were built inside it. Same sequence both times.

The Golden Gate Bridge drawn as a line icon — deck, towers, main cable and suspenders assembling in sequence, the way a harness is built before the load arrives

Do this next

Write the escalation rule this week. If nobody can say where the agent stops and a human takes over, it is not scoped — it is a demo with a roadmap.

Name the one number that proves it worked and put it in the eval harness on day one. If it cannot be measured on a golden set, it is not a metric, it is a hope.

Audit your corpus before you shop for models. Retrieval strategy moved the benchmark more than model choice did. The corpus is the variable you actually control.

Stuck between a pilot and production? Get in touch — we write the inputs first, then build the system that satisfies them.

This is Part 2 of The Forward Deployed Strategist, a four-part series. Next: the economics — who staffs this seat, what it costs, and why the pricing model that built the agency business does not survive contact with agentic delivery.

Powered by Enso Labs

Frequently Asked Questions

What is an agent harness?

Everything in an agent that is not the model — prompt construction, tools, memory and state, permission boundaries, validators, and the evaluation loop. It is the software layer that turns a stateless model into a system that can run a task end to end. Practitioners group it into three tiers: Runtime, Capabilities, and Assurance.

Why do most AI agent pilots fail to reach production?

Not because the model is weak. Deloitte's Tech Trends 2026 finds 38% of organizations piloting agents and only 11% running one in production. Leaders name evaluation gaps (64%), governance friction (57%), and model reliability (51%) as the blockers — all three are harness problems, and all three trace back to inputs nobody defined before the build started.

How does a strategy deliverable become part of an agent?

Directly and one-to-one. The brief becomes the system prompt and the permission boundary. The journey map becomes the tool allowlist and the escalation gate. The measurement plan becomes the eval harness and its golden-set CI gates. The segmentation becomes the retrieval strategy and its recall target. Same craft, executable form.

What does Enso Labs do before building an agent?

Enso Labs writes the inputs first — the outcome the agent serves, the actions it may take unattended, the point it must escalate to a human, and the metric that will prove it worked. Only then does the build start. Get in touch at https://ensolabs.ai/contact.

Want to scope an engagement around this?

Send a brief More insights