Give Them the Harness
A six-week story worth stopping for
In November 2025, Andrej Karpathy wrote about 80% of his code by hand. By December, his agents wrote 80% of it. He called it the biggest change to his workflow in two decades — and it took roughly six weeks. The thing that changed wasn't the model; the model barely moved. What changed was the harness around it. It's the same lesson we keep landing on at BlindOracle, and the one we argued from the cost side in our benchmark on model spend: the scaffolding around the model decides the outcome more than the model does.
Most companies are still hiring for a job that no longer exists at their own company. They write "AI-native" into the job description, put a working agent in front of candidates during the interview — and then, on day one, hand the new hire a stock laptop and a wiki. The harness vanishes the moment the offer is signed.
This piece makes one argument in two halves: provision — give every employee, starting at the interview, a real agent harness; and integrate — stop pretending people don't already bring their own. The governance half of that argument — what you owe once you do it — is its own companion piece: the agent you give your new hire is an identity you've never governed.
The unit of productivity changed under you
For thirty years the unit of knowledge work was a person. You hired the person, measured the person, promoted the person. That unit just got replaced. The new unit is person + harness, and the two are no longer separable. Microsoft's 2025 Work Trend Index gave it a name — the "Frontier Firm," run by teams of humans and agents, where every employee becomes an "agent boss." You cannot make someone a boss and refuse to give them a team. The harness is the team — and once those teammates start transacting with the outside world, they need the same verifiable agent identity any other actor in your systems carries.
| Signal | Number | Source |
|---|---|---|
| Developers using AI tools in their workflow | 80% | Stack Overflow 2025 |
| Code written by Claude at Anthropic | 70–90% | Amodei, 2025–26 |
| PRs shipped in one day by Claude Code's lead (100% agent-written) | 22 | Fortune, Jan 2026 |
| Leaders expecting "digital labor" to expand the workforce in 12–18 mo | 82% | Microsoft WTI 2025 |
What an "agent harness" actually is
An agent harness is the software wrapped around a language model that turns a chat box into a coworker. The field has converged on a clean formula: Agent = Model + Harness. The model reasons; the harness does everything else — tool execution, context and memory, sandboxing, permissions, error recovery, and the project-specific instructions that tell the agent how your shop works. Claude Code, Cursor, Codex, Copilot. The model is a commodity you rent; the harness is the asset you own. The same discipline underwrites our own agent-audit methodology — you can't reason about an agent's output without reasoning about the harness that produced it.
As the practitioners put it: writing code is cheap now; verification is the expensive part. The harness is where verification lives — the tests, the sandboxes, the review loops. It's the same reason a per-call agent economy needs a receipt for every unit of work, the argument we made in why power users are cancelling their AI subscriptions and built into the auditable proof chains behind every BlindOracle deliverable.
You already interview with the harness — then you take it away
The interview is the first place the contradiction shows. The whiteboard is dead as a signal, and the best companies know it:
| Company | What changed | What they grade now |
|---|---|---|
| Meta | "AI-Enabled Interviews" — candidates use real models live | "Critical verification" — catching AI-introduced errors |
| Canva | Scrapped CS-fundamentals; mandatory AI-assisted coding | Judgment on problems you can't one-prompt |
| Shopify | AI embedded in live interviews; ~90–95% of the task | AI judgment — when to intervene |
If the interview assumes an agent harness, the job does too. A company that screens for "critical verification" against a working agent — then onboards the winner onto a bare laptop with a six-week IT queue — has hired for a job that doesn't exist at its own company yet. The signal you selected for is the capability you then strip away. The fix isn't slower onboarding; it's a configured harness on day one, the same way the marketplace onboarding flow hands a calling agent everything it needs to transact from its first request.
Half two: they already brought their own
Provisioning is necessary but not sufficient, because your people didn't wait. "Bring your own AI" isn't a future policy debate — 78% of AI users already bring their own tools to work, 67% of workplace AI interactions happen on personal accounts IT can't see, and over half won't admit to using AI on their most important tasks. The choice was never "allow it or not." Three out of four already did. The real choice is: govern the agents people brought, or stay blind to them — the same trust-an-agent-you've-never-met problem we unpack in trust an agent you've never met.
What "give them the harness" looks like in practice
- Provision at the offer, not the 90-day mark. Full agent fluency takes ~11 weeks and most people quit evaluating in week one. A day-one harness compresses ramp.
- Ship the scaffolding, not just the model. The 34 points live in your instruction files, internal tools, and verify loop. Build once, hand to everyone — the same way agents that pay agents share a common settlement rail.
- Make the standard harness extensible. Let strong hires bring their stack into your perimeter. Good provisioning makes the sanctioned path the better path.
- Interview for the job you actually offer. If you grade verification, build a harness where verification is first-class — the same standard behind our self-audit report.
- Govern every provisioned agent as an identity from minute one. Passport, delegation proof, and audit rail — see the evidence kit and the full governance companion.
Sources (verify primary links before re-quoting): Karpathy 2025 LLM Year in Review · Fowler/Böckeler — Harness Engineering · Cursor 46% vs 80% (via MindStudio) · Stack Overflow 2025 survey · AI interviews — Meta/Canva/Shopify · IBM Cost of a Data Breach 2025 · Microsoft 2025 Work Trend Index. The "bare laptop" line is original framing, not a quoted exec.