How agent harnesses work (a builder's sketch)

Sample article. Template demonstration for builders. Swap this file for a reported guide. Links below are to public papers and vendor engineering posts.

Why this word showed up

Once teams could call tools from a language model, they discovered the model is not the product. The product is the harness: the loop that proposes an action, executes it, observes the result, and decides whether to continue.

Public patterns (ReAct-style thought-then-act, planner/worker splits, computer-use loops) differ in packing, not in the core job.

The loop, without mystique

A minimal harness looks like this:

1. Load the goal, the tool list, and any memory you are willing to trust.
2. Ask the model for the next action (a tool call, a plan, or a final answer).
3. Run that action in a sandbox you control.
4. Append the observation.
5. Stop on success, on a budget, on a policy violation, or when a human takes over.

Everything controversial lives in step 3 and 5: filesystem access, browsers, payments, email. The directory lists sample frameworks that implement variations of this loop.

Memory is a product decision

“Memory” is not one feature. It is at least three:

• Working context inside the context window.
• Retrieved notes (RAG over docs or prior traces).
• Durable state (tickets, carts, git branches) owned by your app, not the model.

Mixing them up is how agents repeat themselves or leak data. Keep durable state in systems you already audit.

Evaluation before aesthetics

A pretty trace UI does not tell you if the agent is useful. Pair every harness with:

• a frozen set of tasks,
• an explicit benchmark or an internal suite,
• a human review sample,
• cost and latency budgets.

See the benchmark leaderboard stub and the sample research note on reading agent benchmarks.

A small reading list

• Our sample breakdown of ReAct.
• The harness comparison table (structure only in v1).
• Glossary: tool use, multi-agent, autonomy.