The model is only part of the story: what an agent harness is

When people compare AI tools, they usually compare models: which one writes better, which one reasons further. But once you hand AI real work, such as a monthly report, a competitor scan or a folder of files to sort, the result depends just as much on the software around the model. That software is called the harness.

The model is the brain, the harness is the workplace

A language model on its own takes text in and gives text out. It cannot open a file, check a figure or remember what you agreed last week. The Claude Code documentation describes the harness as the layer around the model that provides the tools and manages the context the model sees. Anthropic’s engineers put it even more briefly: the harness is the loop that calls the model and routes its tool calls to the right infrastructure.

Think of a brilliant new colleague. Their talent is the model. Their desk, their access badge, the files they may open, the notes they keep and the manager who checks their work: that is the harness.

What a harness does

Most harnesses combine five jobs.

1. The loop. An agent works in cycles. Anthropic describes them as gather context, take action, verify the work, repeat. In its guide Building effective agents, Anthropic calls agents “typically just LLMs using tools based on environmental feedback in a loop”. The harness runs that loop and decides when it stops, for example after a maximum number of turns or at a spending limit.
2. Tools. Tools are the hands: reading files, searching the web, querying a database, sending an email. Every result feeds the next decision. Anthropic writes that while building its coding agent for the SWE-bench benchmark, the team spent more time optimising the tools than the overall prompt.
3. Context and memory. A model can only work with what is in its context window, and that space is limited. Anthropic calls context “a finite resource with diminishing marginal returns”: the more you pile in, the worse the model recalls any single piece. A good harness curates. It loads knowledge when it is needed, summarises older conversation when the window fills up, keeps notes outside the window and hands subtasks to subagents that start with a clean slate.
4. Permissions. What may the agent do without asking? In Claude Code you choose a mode: it asks before every file edit or command, it only plans and proposes, or it works more independently while risky actions are blocked. On top of that, you can allow or forbid specific tools.
5. Human checkpoints. Anthropic recommends that agents pause for human feedback at checkpoints or when they get stuck. OpenAI’s practical guide to building agents names two moments to hand over to a person: when the agent fails too often, and before high-risk actions that are sensitive, irreversible or high-stakes.

Same model, different results

This is not theory. When Anthropic reported its results on SWE-bench, a benchmark built from real software issues, it stressed that the test evaluates an entire agent system, not the model in isolation, and that scores can vary significantly with the scaffolding around it, even with the same underlying model.

A second example comes from Anthropic’s work on long-running agents. Every new session starts with no memory of the previous one. The fix was not a better model but a better harness: a first session that sets up a progress file, a feature list and a version history, so that each later session can see where the work stands. Giving the agent browser tools to test its work as a user would improved results dramatically.

The question is not only which model you use, but what you give it to work with.

And the fit changes over time. Anthropic notes that harnesses encode assumptions about what a model cannot do on its own, and that those assumptions go stale as models improve. A workaround one model needed became dead weight with the next.

What to look for

Whether you are choosing a tool or building your own agent, ask these questions.

• Tools: can it reach the data and systems your work actually needs? A few clear tools beat a long list of overlapping ones; Anthropic names bloated tool sets as one of the most common failures it sees.
• Memory: what does it remember between sessions, where is that stored, and can you see and correct it?
• Permissions: what can it do without asking, and can you set that per person and per tool?
• Verification: does it check its own work, with tests, screenshots or a second review? Anthropic writes that agents that can check and improve their own output are fundamentally more reliable.
• Brakes: can you interrupt it, and does it stop by itself at a turn or budget limit?
• Traceability: is there a log of what it did, and why?
• Model choice: can you move to a newer model without rebuilding everything?

Start with one task you repeat every month. Write down which data, rules and checks a new colleague would need to do it well. That list is your harness.

Zaela is not another chatbot. It works inside the AI you already use, such as Claude, ChatGPT or any assistant that supports MCP, and brings memory, agents, skills and tools together in one place. Your AI sees exactly what you may see, every action is logged, and budgets and limits are set per person.