What it is
“The harness” is a term of art in AI engineering for everything that sits between a raw language model and a working, task-completing agent. A model by itself only predicts text one exchange at a time. The harness is the surrounding code that gives it a job: it decides which tools the model can call, feeds it context about the task and environment, parses and routes its outputs, manages what it remembers as work spans multiple steps, enforces permission and safety checks before actions execute, and decides when a task is finished. A widely cited glossary from Hugging Face researchers frames the harness as “the execution layer” that “calls the model, handles its tool calls, decides when to stop” — distinct from the “scaffolding” of system prompts and instructions that shapes what the model sees and how it should behave. Put simply: the model supplies raw reasoning; the harness supplies the operational discipline that makes that reasoning useful and safe to act on.
Anthropic’s own engineering team, in a November 2025 write-up on building harnesses for long-running agents, describes the concept in practical terms: a harness is what lets a model keep making progress across sessions spanning hours or days, through context compaction, structured file-based memory, and coordination between specialized sub-agents. Anthropic’s Claude Agent SDK is explicitly built and marketed as a general-purpose harness — the company is not just building models but also the control layer around them. Commercial coding-agent products such as Claude Code, OpenAI’s Codex, and Cursor are, in this framing, harnesses wrapped around underlying models, and much of their differentiation from competitors comes from harness design choices rather than from the model itself.
Why it matters for AI governance and narratives
The harness concept has become central to a specific narrative contest: whether the durable commercial and strategic value in AI accrues to the foundation-model labs or to the orchestration layer built on top of them. Perplexity CEO Aravind Srinivas made this argument explicitly to CNBC in July 2026, saying “the model alone is no longer the product” and that the product is instead “the harness, the orchestration system that puts the model inside a very capable harness and pairs the model with a lot of tools.” That framing is strategically convenient for a company like Perplexity, which does not train frontier foundation models itself but builds products atop others’ models — a point the observatory’s symmetric-skepticism principle would flag: this is a claim from an actor with an obvious incentive to devalue the layer it doesn’t own and elevate the layer it does.
The framing also matters for governance and accountability debates. If value and capability increasingly live in the harness — which tools an agent can call, what guardrails constrain it, what data it can access — then regulatory and safety scrutiny aimed only at the underlying model may miss where the actual behavior of a deployed system is determined. Harness design is where permissioning, approval gates, and revocable identities (the kind of agent-identity infrastructure companies like Better Auth are building) actually get enforced. As foundation models converge in raw capability, competitive and governance attention is following the harness — a shift the observatory has already flagged in coverage of the Linux Foundation’s new Agentic AI Foundation and moves toward scoped, revocable agent identities.
Key facts and dates
The term has proliferated rapidly through 2026 as agentic AI products matured. Hugging Face researchers Sergio Paniego and Aritra Roy Gosthipaty published a glossary attempting to standardize the vocabulary (“Harness, Scaffold, and the AI Agent Terms Worth Getting Right”) on May 25, 2026, motivated by inconsistent industry usage — some products use “harness” loosely to mean everything except the model itself. Anthropic published detailed engineering guidance on harness design for long-running agents on November 26, 2025, and has continued publishing on the topic through 2026, treating the Claude Agent SDK as its flagship harness product. Academic attention has followed: a March 2026 arXiv paper by Nghi D. Q. Bui examines harness, scaffolding, and context-engineering design specifically for terminal-based coding agents, treating these as the practical engineering levers that determine whether an agent succeeds at real-world tasks. Aravind Srinivas’s July 2026 CNBC remarks are the most direct evidence of the term migrating from engineering jargon into public industry strategy discourse — evidence the observatory’s editorial cited directly.
Where to learn more
- Anthropic: Effective harnesses for long-running agents — primary source, Anthropic’s own engineering framing of the concept
- CNBC: The AI race is shifting from bigger models to cheaper, smarter systems — source of the Perplexity CEO quote referenced in the editorial
- Hugging Face: Harness, Scaffold, and the AI Agent Terms Worth Getting Right — attempt at a standardized technical glossary, May 2026
- arXiv: Building Effective AI Coding Agents for the Terminal — Scaffolding, Harness, Context Engineering, and Lessons Learned — academic treatment, March 2026