The harness is the half you own

A year of model launches trains a habit: read the benchmark column, note that the frontier moved again, wait for the next one. The column is real, and we read it the day it ships. But the column is not what an operating business runs. What it runs is an agent, and an agent is not a model.

The equation is almost insultingly simple. model + harness == agent. The phrasing is Birgitta Böckeler’s, writing for Thoughtworks: the word harness has become shorthand for everything in an agent except the model itself. The model is the frontier weights, the same ones your competitor can buy, improving on a clock you do not set. The harness is everything else, and it is the only half you build.

agent == model + harness (the decomposition, after Böckeler)
  model     the frontier weights. the same ones your competitor can buy,
            improving on a clock you do not set. not yours to change.
  harness   everything else: tools, context, memory, feedback loops,
            guides and sensors, orchestration. the half you build and own.
  agent     where model and harness meet: the thing you actually run.

Figure 1. an agent is a model plus a harness, and the split is Böckeler’s. the model is the frontier weights, bought and improving on its own; the harness is everything else (the tools, the context, the memory, the feedback loops, the guides and sensors, the orchestration), the half you build. where the two meet, and only there, sits the agent: the thing you actually run.

What the harness is

Strip the mystique and the harness is plumbing. It is the tools the model is allowed to call and the ones it is not. It is the context it sees and the order it sees it in. It is the memory that survives between turns, the feedback loops that tell it whether the last step worked, the orchestration that breaks a long job into pieces it can finish. Böckeler splits the builder’s half into two moves: guides, which steer the model before it acts, and sensors, which watch after it acts and let it correct. OpenAI, describing five months of building software with coding agents and almost no hand-written code, put the same point plainly: the discipline now shows up in the scaffolding rather than the code.

Seen running, the harness is a loop. A prompt goes in, and the model works inside a cycle the harness defines: gather the context it needs, take an action, check whether the action worked, and go round again until the job is done. The model supplies the judgment at each step. The loop, the tools it may reach for, and the checks that catch its mistakes are all harness, and a person can lean in at any turn to steer.

the agentic loop (the harness running around the model)
  your prompt
    -> gather context -> take action -> verify results -> repeat
    the loop, the tools it may call, and the checks it runs are the harness;
    the model only acts inside it
  -> done
  a person can interrupt, steer, or add context at any turn

Figure 2. the harness as a loop. a prompt goes in; inside the agentic loop the model gathers context, takes an action, and verifies the result, repeating until the work is done. the loop itself, the tools it is allowed to call, and the checks that catch its mistakes are the harness. the model only acts inside it, and a person can interrupt, steer, or add context at any turn. after Anthropic’s account of long-running agent harnesses.

Everything we have shipped is a harness in this sense. The appointment agent is a model with every consequential action gated in code. The voice agent is a model wrapped in a transport that listens, offers a slot, and reads it back. The model is bought. The agent is built.

The harness ages; the model improves

Here is the tension the word hides. The model improves on its clock. The harness sits still unless someone moves it. And every piece of a harness, as Anthropic’s Prithvi Rajasekaran writes, encodes an assumption about something the model could not do on its own. Those assumptions are exactly the things that go stale. The workaround you wrote because last year’s model lost the thread on long tasks, the extra validation step because it used to invent a field, the decomposition because it could not hold the whole job at once: each was true when you wrote it, and each becomes dead weight the moment a better model no longer needs it.

A harness that ossifies around an old model’s deficits does not merely stop helping. It starts to subtract. It holds a frontier model down to the shape of the one it was built for. When Anthropic upgraded the model under one of their own harnesses, they could delete an entire scaffolding construct the new model had made redundant, and the work got better, not worse. The useful boundary, in their words, moved outward.

That is the whole of the thesis. A harness is not timeless because it is finished. It is timeless because it does not get in the model’s way. The durable harness is the one whose leverage moves outward as the model improves, so each capability gain flows straight through to the business rather than breaking against scaffolding built for a weaker model.

model + harness == agent (schematic, not a measurement)
  axis x   successive frontier models, each better than the last
  axis y   what the agent actually delivers to the business
  frontier capability          rises on a clock; the model is not yours to change
  a harness that moves with it  tracks the frontier; each gain flows through
  a harness that ossifies       encodes the old model's deficits; it peaks, then
                                fights every better model and falls behind
  the widening gap between the frontier and the ossified harness is
  self-inflicted: the scaffolding, not the model, is the ceiling

Figure 3. the same model behind two harnesses. the one that moves with the model tracks the frontier; the one that ossifies around an old model’s deficits peaks, then falls away as each better model fights its stale scaffolding. the gap that opens is not the model’s fault. this is a schematic, not a measurement: it names the shape of the argument, not a measured curve.

Timeless is not thin

It is tempting to read this as an argument for a minimal harness, a thin shim that does as little as possible and lets the raw model shine. That reading is wrong, and the same sources that support the thesis refute the shortcut. Rajasekaran is blunt that the space of interesting harness combinations does not shrink as models improve. OpenAI’s lesson from a million lines of agent-written code is that effort moved into the scaffolding, not out of it: their harness is heavier than what it replaced, not lighter, and deliberately model-native, shaped to how the model actually performs best so its capability is unlocked rather than boxed in.

The resolution is that timeless and minimal are different words. A harness gets in the model’s way not by being large but by being stale, by encoding and then defending assumptions the model has outgrown. The rule is not build less. It is build nothing the model will soon do better, and stand ready to remove what it already does. Find the simplest thing that works, Rajasekaran writes, and add complexity only when it earns its place. The harness can grow rich. It cannot be allowed to set.

The frontier is coming for the harness too

The clearest sign that the harness is now the real engineering surface is that the frontier has begun to automate it. A Stanford group, the IRIS Lab, recently published Meta-Harness, a system that puts one agent in charge of writing and rewriting the harness code for another, searching over scaffolding the way earlier systems searched over prompts. The detail that matters: it reads the raw execution traces of every prior attempt, not just the scores, because the score throws away most of what went wrong. On Terminal-Bench, the hard command-line benchmark from Stanford and the Laude Institute, the authors report their discovered harnesses beating the best hand-built ones on the same model. That figure is the authors’ own, not yet an independent leaderboard result, and we flag it as such.

Read it for the direction, not the number. Even harness-building is becoming something you maintain rather than finish, an evolving thing kept in step with the model under it. The hand-built harness that is never revisited is already a legacy system.

The half you own

For a lab whose whole practice is wiring frontier models into operating businesses, this is not an abstraction. It is the job description. We do not improve the model. That is Anthropic’s work, and OpenAI’s, and it happens on a clock no client controls. We build the harness: the half of the agent that sits inside your operation, the half that either compounds with the frontier or fights it. We argued in the capability overhang that closing the distance between what AI can do and what your business does with it is embedding work, not tooling. This is the same claim seen from the inside. The embedding is a harness, and a harness built well is one that gets quietly better every time the model behind it does, without anyone touching it.

If there is an agent running somewhere in your operation that has not improved since the day it was wired up, the model is almost certainly not the reason. Start a conversation with us about a Discovery Phase.

References

Birgitta Böckeler, Thoughtworks. Harness engineering for coding agent users. 02 Apr 2026. martinfowler.com/articles/harness-engineering.html
Anthropic. Harness design for long-running application development. Prithvi Rajasekaran, 24 Mar 2026. anthropic.com/engineering/harness-design-long-running-apps
OpenAI. Harness engineering: leveraging Codex in an agent-first world. 11 Feb 2026. openai.com/index/harness-engineering
OpenAI. The next evolution of the Agents SDK. 15 Apr 2026. openai.com/index/the-next-evolution-of-the-agents-sdk
Stanford IRIS Lab. Meta-Harness: End-to-End Optimization of Model Harnesses. arXiv:2603.28052, 30 Mar 2026. arxiv.org/abs/2603.28052
Stanford University and the Laude Institute. Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces. arXiv:2601.11868, 17 Jan 2026. arxiv.org/abs/2601.11868