In Harness

Part 6 of 7 in Atoms to Applications, Mark Pesce's walk through the AI stack. First published by Mark Pesce in The Watershed on 2026-10-13. Republished with permission.

A harness turns a model into something useful. A model by itself is simply a next-token-predictor; inside of a harness it can plan, execute, and achieve goals set by its user.

That makes harnesses incredibly important, and a core focus of Noops Research as we explore the 'Atoms to Applications' stack in preparation for our AI Investor event.

As far back as the first version of ChatGPT (30 November 2022), we've used harnesses. A chatbot is a harness because it maintains conversational state across the context window of the model, feeding the entire conversation back into the model at every turn; the model, by itself, is largely stateless.

The evolution of the chatbot is the evolution of the harness: First text, then files, then images, then project folders, then skills, then CLAUDE.md and AGENTS.md and so on. There's no clear line separating a chatbot from a harness. Anthropic folded Claude Cowork - a general-purpose harness - into Claude chat in September 2026, making it official.

Nevertheless, the recognition of a harness as a harness is an identifiable event - and it comes with release of Claude Code in 2025.

Agents are specific instances generated by harnesses, and inherit all the qualities of the harness that creates them. A harness can create an "infinite" supply of agents - to the computational limits of the model within the harness.

Coding Harnesses

In February 2025, Anthropic released Claude Code, a terminal-based harness for software engineers. Although neither Claude nor Claude Code were particularly good at that moment in time, the combination of model + harness was "good enough" to ignite an explosive release of creativity, both among software engineers using these tools and among software engineers building similar tools.

Andrej Karpathy coined the term 'vibe coding' to describe the new capability to direct a harness to 'do this thing', and have it done more or less correctly. From February through November 2025 both models and harnesses grew much more capable, leading to a 'watershed' moment (sighted first by Simon Willison) coincident with the release of Gemini 3, Opus 4.5 and GPT 5.2, when software engineers stopped 'fighting' the harness and started running to keep up with their high-quality outputs.

By the end of 2025, coding harnesses included OpenAI Codex, Gemini CLI, Cursor, Windsurf, OpenCode and Pi. Windsurf was acqui-hired by Google to become Google Antigravity, while Cursor was purchased for US$60,000,000,000 by SpaceX @shortly after their 2026 IPO.

Why pay so much for a harness?

Harnesses Matter

A bad harness makes a good model stupid. A great harness makes a stupid model smarter.

Testing revealed OpenAI's GPT-5.5 scored 61.5% in Codex and 87.2% in Cursor, a 25.7-point swing on an identical model - far larger than the gap between most model generations. OpenAI's model did best in a third-party harness.

Testing published in September 2026 covered nine harnesses using the same model (Kimi K3), with same hardware, across 360 trials. Pass rates ranged from a high of 66.7% with OpenAI's Codex to 50.0% with both OpenCode and Hermes. Median cost per task ranged from a low of $1.05 with the Exo harness, to $18.34 using Claude Code, a 17x cost spread - due to the harness.

Any organisation deploying agents at scale has to consider its harness very carefully, balancing its use cases against the strengths and weaknesses of the harness.

General-Purpose Harnesses

In November 2025 - in sync with the Watershed - Peter Steinberger released the first version of a general-purpose agent that would eventually be known as 'OpenClaw'.

A general-purpose agent differs from a software agent in that it has a much richer set of tools to execute the goals set by its user. OpenClaw has a strong standard tool library, with another nearly 5000 tools available through hubs such as ClawHub, enabling integrations with many common software packages, such as Google Workspace, Obsidian and Notion.

With its open architecture, its broad capability, and its friendly attitude - supported by a SOUL.md file that gave it a breezy, humorous tone - OpenClaw became an overnight sensation - purportedly so popular it generated a 'run' on Apple MacMinis that could run OpenClaw securely on an isolated piece of hardware.

Not even a year later, OpenClaw seems to have faded from memory; its candle burned too brightly, and too early, exposing the security problems that make general-purpose agents both powerful and potentially dangerous. When a general-purpose harness has access to everything, small errors can cascade into big problems.

NVIDIA created a security-aware fork of OpenClaw, known as NemoClaw, which has seen some adoption in corporate environments. Corporate-scale adoption of general-purpose agents will be gated by security concerns, and by CISOs who will not allow such unconstrained agents anywhere near corporate networks.

Domain-Specific Harnesses

There are as many potential harnesses as there are operational domains. Here are some of the standout examples:

• Design: Figma Design Agent (closed beta, 20 May 2026). Acts on the canvas, reads and writes layers, components, tokens and variables, hands off to Figma Make and the Figma MCP server. Memory is the design system itself plus file history.
• Engineering/CAD: Zookeeper in Zoo Design Studio (Jan 2026, on Text-to-CAD from Dec 2023). Generates KCL, executes against Zoo's B-Rep engine, runs mass and volume analysis, reads its own docs; explicit plan-act-observe-update loop with geometry compile as verifier.
• Biology: Kosmos (Edison Scientific, spun out of FutureHouse; 5 Nov 2025, arXiv 2511.02824; Nvidia version March 2026). A literature agent over ~175M papers plus a data-analysis agent that writes and runs code, sharing a structured "world model" as memory across hundreds of trajectories. Up to 12 search-hypothesise-analyse-update cycles per run; every conclusion traceable to code lines or passages. One run reads ~1,500 papers and writes ~42,000 lines of code; 79.4% of statements judged accurate by experts; 7 validated discoveries (3 reproductions, 4 novel).
• Chemistry/Materials science: Periodic Labs (Fedus and Cubuk; $300M seed at $1.3B Sep 2025, raising ~$500M at ~$7.5B May 2026, Nvidia among backers). Robotic powder synthesis, robot-loaded XRD, quantum simulation, models reading diffraction patterns to choose the next experiment; a proprietary dataset including failed syntheses used to mid-train the models. ~100 experiments a day, scaling to 1,000; early semiconductor-industry customers.
• Law: Harvey ($200M Series C at $11B, 25 Mar 2026). "Tool Bundles" (file-system bundle with grep-like and semantic search over the matter; drafting bundle with a sub-agent), Ask LexisNexis, iManage, Word; customers have built 25,000+ custom agents. Memory persists preferences across sessions plus matter-scoped context; Harvey explicitly manages "context rot." A plain agentic loop, deliberately avoiding hand-written workflow orchestration, with leave-one-out validation gating per Tool Bundle. 100,000+ lawyers, 1,300+ organisations, ~$190M ARR (third-party estimate).
• Finance: Bridgewater Associates Labs (testing from late 2023, outside investors July 2024). Causal time-series models, frontier and specialised LLMs, century-scale proprietary macro databases, an AIA Forecaster matching human superforecasters, a Pocket Analyst running research agentically, wrapped in "systems designed to test and diagnose their outputs." Memory is Bridgewater's codified cause-and-effect investment logic plus live trade feedback. Loop: gather, explain causally, test, position, let the market grade it.
• Medicine: MAI-DxO ( Microsoft AI, 30 Jun 2025, arXiv 2506.22405). A model-agnostic orchestrator running a virtual panel (hypothesis, test selection, challenger, cost steward, checklist) over any frontier model; keeps the evolving case record and running cost; orders the next most informative test before committing. 85.5% diagnostic accuracy with o3 against 20% for 21 experienced physicians, at ~20% lower test cost, and it lifts every model it wraps.
• Semiconductor design: Cadence ChipStack AI Super Agent (10 Feb 2026; Cadence and Nvidia "fully autonomous EDA agent" 1 Jun 2026). Drives Verisium, Cerebrus, JedAI and Xcelium-class simulators with NeMo and Nemotron underneath; generates RTL and testbenches, writes test plans, runs regressions, debugs and auto-fixes; a "Mental Model" grounded in the spec as memory. Claimed 10x front-end productivity; Nvidia and Altera among 10+ early-access customers.

The number of operational domains is vast, and will grow as harnesses grow more specific to a particular subdomain. Every user of a harness will be looking for the harness that provides the greatest 'lift' for their task.

Self-Improving Harnesses

Rather than improving the model weights - the frontier labs' approach to recursive self improvement (RSI), more might be gained by applying the same principles to harness design. Lillian Weng has written an excellent post on this topic.

• Self-Harness (23 Jun, arXiv 2606.09498): agent mines its own traces, proposes minimal harness edits, keeps only those passing regression tests; Terminal-Bench 2.0 held-out pass from 40.5% to 61.9%.
• HALO (24 Jun, context-labs): the production version, OpenTelemetry traces in, harness edits out, redeploy, repeat.
• AutoSaddler (30 Aug, arXiv 2608.23041, 13 authors): treats harness improvement as offline learning, treating the harness as code an optimiser can patch; +9.0 on GAIA2, +9.6 on SWE-Bench Pro, +10.0 on Terminal-Bench 2.0.

The human harness-engineer leaves the inner loop and per-deployment integration cost collapses. As harnesses self-improve, the moat in hand-built domain harnesses decays toward proprietary state and distribution, which is an echo of the Bitter Lesson.

The Future

The harness is the current coalface of development in artificial intelligence; now that models are 'good enough', we need harnesses that can lift them into specific capabilities. These harnesses are likely to come from within rather than outside of the operational domains they are designed for. "Forward Deployed Engineers" are likely to spend most of their time helping customers build their own harnesses.

Atoms to Applications, part 6 of 7 · ← This Year's Model · Growing Applications →