Skip to content
Agentic Control Plane

What Is an Agent Harness? (And Why It Needs a Control Plane)

David Crowe David Crowe · · 6 min read
harness agents governance architecture
Share X HN LinkedIn

An agent harness is everything around the model that turns it into a working agent: the loop that runs it, the tools it can call, the context and memory fed into each turn, the checks, the budgets and the logging. The model decides. The harness acts, and most of the cost, reliability and risk sits there.

A line going around names it: Agent = Model + Harness. The model gets the attention. The harness is most of the system.

An agent is mostly harness Agent = Model decides + HARNESS — acts, and it's most of the system Loop Tools Context & memory Verification Budgets & limits Logging The model decides what to do. The harness does it, and carries the cost, the risk and the surprises.

The model is the smallest part

Give a frontier model a prompt and it returns text. That isn’t an agent. An agent reads a ticket, searches the web, queries a database, drafts a reply and sends it. It loops, checks its own work and decides what to do next.

The model does none of that by itself. The harness is:

  • the loop that runs the model again and again until the task is done,
  • the tools it can call, and how their results go back in,
  • the context and memory: what gets loaded into each turn and what gets trimmed,
  • the verification: checks that run before a result reaches a user,
  • the budgets and limits: how much a run may spend and how long it may go,
  • and the logging of what happened.

Put a better model in a bad harness and you still get a bad agent. Fix the harness and keep the model, and the same model ships. The gap between what a model can do and what you see it do is called the harness gap. Closing it is harness engineering, now a discipline of its own. Thoughtworks, Databricks, LangChain and MongoDB have all written about it.

Cost and surprises come from the harness

Once you see an agent as a harness, odd behavior makes sense.

If a run costs more one day than the next, it is usually not a flaky model. The loop ran a different number of turns. A bill that balloons on a simple task often means the harness re-sent its whole context on every turn. And a run that takes three paths to the same goal is the loop diverging.

These are measurable: turns, how often context is re-read, path variance, cost per step. Harness engineering makes an agent’s behavior something you can see and tune.

One part of the harness has its own pattern language, the control layer: what the agent may do, as opposed to what it can do. We catalogued eight patterns that harnesses keep arriving at (prompt rules, permission prompts, pattern rules, approval classifiers, hooks, sandboxes, proxies and independent ledgers) in harness control patterns.

What the harness can’t solve alone

Once a harness calls real tools, spends real money and acts for real users, five problems show up at the point where it touches your backends:

  • Identity. When the harness calls your API, who is it acting as? The user who triggered it, the agent, or the framework’s service account? Most harnesses paper over this with one shared key. This is the three-party problem.
  • Authority. What may it do? Delete, pay, email the whole company? A harness will call any tool you give it and has no view on which calls are dangerous.
  • Cost. What is it spending, per agent and per tool call, and can you cap it before a runaway loop bills you hundreds of dollars?
  • Control. Can you hold one risky call, or block one tool, without a redeploy?
  • Evidence. Can you show an auditor every action, as a specific user, with the result?

You can build a good harness and still have all five. They apply to every tool call, whichever framework you used.

Observability and governance are different jobs. Observability tells you what the harness did. Governance decides what it’s allowed to do. Traces and logs are read-only. Identity, rules, budgets and blocks are enforcement. A harness in production needs both.

What a control plane does here

A control plane sits at the tool-call boundary and covers those five things. This is also where the harness conversation keeps ending up: the AI control plane.

ACP’s version, for coding agents:

  • A hook in the harness sends each tool call to ACP before it runs, along with who is asking.
  • Your rules answer allow, ask or deny, per tool. A safety floor holds force-pushes, destructive SQL through a database client, curl | sh and deletes outside the project for your approval, and always blocks things like rm -rf of your home directory.
  • Every call is logged with the user and the decision, and model calls are priced where they go through ACP.
  • A spending cap in the console can stop tool calls when it’s reached.

A new workspace starts in audit mode: everything is recorded, and only the floor holds anything. You add rules after you’ve seen what your agents do.

Limits worth stating. This is a hook, so it covers the harnesses that have one. See the coverage table for which harnesses, which tools and how a hold reaches you. Only Claude Code shows a safety-floor hold inline in the session. Everywhere else, and for holds from your own rules, the call is blocked, you approve it in the console or from the email, and the agent retries. The hook runs on the machine where the agent runs, so a session that disables hooks disables it.

We build ACP, which is why we’ve spent so long looking at harnesses. You engineer the harness for capability. Whatever you call the layer under it (control plane, governance layer, agent gateway), a harness that does real work needs one.

Running coding agents? ACP records each tool call its hook sees and makes the dangerous shell commands (a force-push, destructive SQL, curl | sh) wait for you. One command, free up to 5 agents:

macOS · Linux · WSL
curl -fsSL https://agenticcontrolplane.com/install.sh | bash
Windows · PowerShell
irm https://agenticcontrolplane.com/install.ps1 | iex

What the installer writes  ·  Getting started

Share X HN LinkedIn
Get the next data drop
What agents actually cost, new tool-surface captures, and the occasional incident post-mortem — sent when we publish something worth your inbox, not on a schedule. Unsubscribe anytime.
Share: Twitter LinkedIn
Related posts

← back to blog