Understand the Model Harness
How context, tools, permissions, state and evaluation turn a model into an operated product.

On this page
A research assistant cites an old document while answering a question about a current process. The model may have misunderstood the question. It may also have received an outdated source or failed to retrieve the current one. Replacing the model without inspecting the path may leave the problem intact.
The harness is the system around the model that makes the task possible. Product managers do not need to implement every part of it, but they should understand enough to make quality, experience, cost and risk decisions with the people who do.
If models, context or retrieval are unfamiliar, start with how models fit into a product.
Trace a task through the system
| Component | Product question |
|---|---|
| Model | What capabilities and limitations matter for this task? |
| Context | Which current sources, examples and instructions does the system receive? |
| Tools | What information can it retrieve and what actions can it request? |
| Permissions | Whose authority applies, and which actions need an independent check or approval? |
| State | What must be remembered during or between tasks, and for how long? |
| Execution | Where does work run, and what limits apply to time, cost and access? |
| Evidence | How can the team inspect results, traces, failures and changes? |
For the hypothetical research assistant, the harness receives a question, checks which sources the user may access, retrieves relevant passages and supplies them to the model. The model prepares an answer; the product displays it with references so the reader can check the source. If the system can also change a record, another layer must validate that action and enforce the user's authority. The model's instruction is not the permission check.
Follow a wrong answer
The user asks for the current leave policy and gets a confident answer based on last year's document. The team can inspect the recorded path, or trace, subject to the user's access and the trace's privacy controls:
- Source: Is the current policy actually available in the approved collection? If the collection itself is stale, identify who maintains it and fix the update process before tuning the model.
- Retrieval: If the current policy is available, did the search return it? An outdated result may point to indexing, access, ranking or the wording of the query. Compare the retrieved passages and their dates.
- Model and instructions: If the current policy reached the model, did the answer use it correctly? Test the same case again with the relevant context visible. A wrong interpretation is a reason to inspect the instruction and model behaviour, though one run cannot isolate the cause conclusively.
- Presentation and recovery: Did the interface expose the date and source, and could the user report or correct the error? Even a retrieval problem becomes harder to catch if the answer hides its evidence.
These checks narrow the investigation; they do not prove a single cause. Work with the engineers and policy owner, change one part where practical and rerun the affected cases. If traces are unavailable or omit the relevant input, add appropriate observability before drawing a confident conclusion.
Treat context as a maintained dependency
Use relevant information from sources that someone maintains. Adding more context can introduce stale rules, conflicting examples and unnecessary data. Keep track of where information came from when a user or operator needs to verify it.
Separate task state from durable preferences, shared knowledge and interaction history. Each has different access, retention and correction needs. Test material changes to context and tool behaviour against the affected cases in the evaluation suite.
Give tools a clear contract
For each tool, define the job it supports, the permitted data and action, validated inputs, possible errors and what happens on timeout or partial completion. A retry must account for an action that may already have succeeded. Log enough for an owner to investigate without collecting unnecessary sensitive data.
The description helps the model choose a tool. Independent validation decides whether the requested action is allowed. The security playbook covers untrusted content and authority boundaries.
Decide whether routing helps

A router chooses a model or execution path for a task. Different paths may meet different requirements for quality, latency, availability, data handling or cost. Routing also introduces classification errors, extra calls, maintenance and evaluation work.
Start with a path that meets the requirements. Compare alternatives on representative tasks and include the full cost of context, tools, retries and review. Add a route when it provides a measurable benefit or satisfies a necessary constraint. A team that cannot tell whether routes meet the quality bar is not ready to optimise by price alone.
Keep a clear boundary around model-specific behaviour and record dependencies. Replacing a model may still require changes to prompts, tools and tests; a clean interface does not make models interchangeable. Business viability covers the economic decision, and agent patterns covers when additional agents may help.
Try it: inspect an incomplete trace
Imagine a research assistant answers a question about the current returns process. Its answer cites a document dated last year. The trace shows that a search returned three passages and the model received one of them. It does not show the search query, the other passages or when the source collection was updated.
Before looking for a fix, write down what you know, at least two possible causes and what you would inspect next. For example, the source collection may be stale, or retrieval may have selected an older document. The trace does not establish either explanation. Ask an engineer or the source owner what evidence would distinguish them, then agree on a test and an appropriate way to protect any sensitive trace data.
If you have access to a real trace, repeat the exercise with permission. Check the result with the people who maintain the system. A model can suggest hypotheses from a trace, but it cannot see omitted steps or prove a cause from missing information.
v3.1 · Updated Oct 2026