The Agent Harness
An agent harness is the orchestration layer above the model: the tooling that holds your context, reads your codebase, executes your workflows, calls your tools and enforces your policies. The model supplies raw intelligence. The harness turns it into work — inside your environment, under your rules.
Every agent you use runs inside one. A coding agent that reads your repository, edits files, runs tests and opens pull requests is a harness with a model plugged into it. The model can be swapped — this quarter's flagship, next quarter's cheaper open-weight equivalent — and the user barely notices. What cannot be swapped cheaply is everything the harness has accumulated: the context, the tool integrations, the permissions, the workflow shapes, the history of what worked.
Models improve fast and get cheaper faster. The durable question is not "which model wins" — it is who owns your harness, because that layer owns the switching costs.
This is the standard economics of commoditizing platforms, replayed one layer up. When capability becomes abundant — and capability is commoditizing — value migrates to the scarce layer adjacent to it. In the AI stack that layer is the harness: it is where a task gets routed to the cheapest sufficient model (the Joule Wars lever), where evidence of correct execution gets captured (the Proof-Adjusted Autonomy gates), and where an organization's actual working context lives.
What a harness does
- Holds context. Your codebase, documents, conventions, memory — assembled per task, kept current, kept private.
- Executes. Tool calls, shell access, file edits, API integrations — the hands of the agent, scoped by permissions.
- Routes. Chooses which model runs which task; the cheapest sufficient model by default, the flagship on the hard tail.
- Enforces policy. What the agent may touch, what needs human sign-off, what is irreversible.
- Captures evidence. Logs, diffs, test results — the proof package that makes autonomous output acceptable.
Why ownership matters
Whoever operates the harness sees your workflows, accumulates your context and collects the switching costs. Model labs are pushing down into the layer (Claude Code, Codex CLI); platforms push up from the IDE and the cloud; enterprises quietly assemble internal harnesses around their own data and policies. The strategic audit for any organization is one question: is your harness — and the context inside it — owned, rented, or leaking? The full argument is in Who Owns Your Harness?.
Frequently asked questions
What is an agent harness?
The orchestration layer above the model: the tooling that holds your context, reads your codebase, executes your workflows, calls your tools and enforces your policies. The model supplies raw intelligence; the harness turns it into work inside your environment. Coding agents are harnesses; the model behind them is a swappable component.
Why does the harness matter more than the model?
Models commoditize — open-weight and mid-tier models keep closing the capability gap — while the harness accumulates what doesn't: your context, workflows, permissions and history. Swap the model and users barely notice; swap the harness and you lose everything. The switching costs live in the harness.
Agent harness vs agent framework — what's the difference?
A framework is code you assemble into an agent at build time. A harness is the operating environment a finished agent runs in at run time — context, tools, execution, guardrails, evidence. Frameworks compete for developers; harnesses compete for the workflow, which is where economics and lock-in are decided.
Who owns the agent harness market?
It is being contested now: model labs push down (Claude Code, Codex CLI), platforms push up from the IDE and cloud, enterprises build internal harnesses around their own context. The question that matters for you: is your harness owned, rented, or leaking to a vendor?
How does the harness relate to Joule Wars and PAA?
The harness enforces both frontiers: it routes each task to the cheapest sufficient model (the Joule Wars lever) and captures the evidence that makes agent output acceptable (the Proof-Adjusted Autonomy gates). Efficiency and proof are harness features, not model features.
Go deeper
Definition maintained by Michał Piszczek — CTO of Archdesk, author of the Joule Wars and Proof-Adjusted Autonomy frameworks. Quote freely with attribution to piszczek.pl/agent-harness (CC BY 4.0).
Piszczek, M. The Agent Harness — the orchestration layer above the model (definition). piszczek.pl. https://piszczek.pl/agent-harness