The 2027 Harness Wars: Why No Enterprise System of Record Can Stay Neutral

For two years, the enterprise AI conversation has been a model conversation: which foundation model, which benchmark, which vendor. That framing just took a hit. NVIDIA and LangChain announced that tuning the Deep Agents harness — the prompts, tool definitions, and middleware sitting around a model — let an open-weight Nemotron model reach accuracy parity with the highest-scoring closed models, at roughly a tenth of the cost per task. No retraining, no new weights. Every gain came from engineering the environment the model runs in, not the model. That's a signal the layer everyone ignored — the harness, the agent runtime, the execution environment — has become a lever worth as much as model choice, and in some deployments, worth more.

The term moved just as fast as the idea. Anthropic wrote about harnesses for long-running agents late last year, Mitchell Hashimoto's essay on his own AI adoption helped popularize "harness engineering" as a discipline, and OpenAI published its own take within weeks. Once you accept that the harness decides what an agent can actually reach and do, you've accepted that whoever owns the harness owns the leverage — and that turns a systems-engineering question into a strategy question for every company whose product is, at its core, a place where an organization's data lives.

That's the binary now facing every system of record — CRM, ITSM, ERP, HRIS, core banking, take your pick. Option one: expose an API or an MCP endpoint and let external agents pull data through it. The system becomes a commodity data tap — fungible, interchangeable, one contract renewal away from being routed around by whichever orchestration layer actually wins distribution. Option two: become the harness itself, the place where the agent executes, with the platform supplying not just data but the approvals, guardrails, and audit trail that make execution defensible. One path keeps the data. The other keeps the workflow, the relationship, and the liability that comes with it.

Salesforce picked its side loudly. Agentforce inside Slack now queries CRM records, executes multi-step actions, and routes approvals directly inside the conversation threads where decisions already happen — a four-layer architecture with Slack as the surface, Agentforce as the agent, and Customer 360 as the data underneath. What's notable isn't the feature list; it's that Salesforce is willing to call this a structural shift, not an add-on. And this isn't confined to companies built AI-native from day one. At its May 2026 community forum, Temenos, whose core banking software runs the ledgers of banks with far less appetite for architectural risk than a SaaS company, told its customers they don't need AI "added on top of critical systems" — they need it built into the workflows they already trust. If a 30-year-old core banking platform is willing to make this bet, "our industry is different" stops being a credible reason to sit this one out.

Here's the part risk, legal, and compliance will care about more than the architecture diagram: becoming the harness only pays off if the platform can actually govern what happens inside it. An execution environment that lets agents act without enforced approvals and a real containment layer doesn't capture a moat — it just moves the liability closer to home. That's the same gap the Agentic AI Enterprise Blueprint and Pilots to Platforms pieces flagged around owning the runtime and the evaluation harness, and it's the same distinction the CLEAR framework draws between monitoring a system and being able to stop it — built originally for credit decisioning, but the logic travels to any system now being asked to act autonomously on an organization's behalf.

Call it what it is heading into 2027: a harness war, and not a metaphorical one. Every system of record — yours included, whatever industry you're in — is being forced to choose sides, whether or not its roadmap says so. Companies that treat this as a vendor integration decision will lose the war without realizing they were fighting it. The ones that treat it as a governance decision, with TRUST framework-level discipline about accountability built in from day one, are the ones that will still own their customer relationships when the dust settles.

For why owning your context layer matters in this contest, see the 2027 context layer.