> ## Documentation Index
> Fetch the complete documentation index at: https://docs.within.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Maintain and improve agents

> Keep an agent current as the work changes: record its baseline, watch for process changes, classify each change and retest before you update.

An agent relies on assumptions about the process, its source material, the tools it calls and the controls around it, and all of them change. Your Work Brain keeps learning from the work as people use Companion, so maintenance starts the day the agent ships: compare what the agent was built on with what the work looks like now.

## Record the baseline

Store these with the agent's code and evaluations, so every later review has something to compare against:

* The process it was built from, and the version used.
* The evidence behind its behavior: the observations or an evidence summary.
* The policies, dependencies, tools and model configuration it relies on.
* The evaluation set and the release result.
* The owner, and the next review date.

## Know when to review

* **On a schedule.** Set the next review date in the baseline.
* **When the process changes.** A known change to the process, its owner or its systems.
* **As capture comes in.** Companion needs to run often enough to cover the processes and exceptions the agent depends on.
* **From the change feed.** The Within MCP returns recent process changes. It's a bounded set with no date cutoff, so it doesn't prove nothing changed before its window.

## Classify the change

Start with recent changes, then pull the current details of each process the agent uses and compare them with the baseline. Where behavior matters, look at current observations too: the documented process can stay the same while the way people run it, its exceptions or its systems change.

| Change | Example | What to do |
| - | - | - |
| **Presentation** | A rename that doesn't change the meaning | Update references and rerun the affected evaluations |
| **Behavioral** | A rule, exception, input or handoff changed | Revise the build and rerun the full evaluation set |
| **Structural** | The objective, owner, system boundary or a required approval changed | Go back to the use case and the design before you rebuild |

The classification is your decision, not the tool's. Confirm material changes with the process owner before you change how the agent behaves in production.

```text Maintenance review theme={null}
Compare [agent]'s recorded baseline with the current evidence in Within for the processes it uses. List what changed, classify each change as presentation, behavioral or structural, and propose an update and the evaluations to rerun. Do not modify the agent until I approve the proposed response.
```

## Decide what's next

| Decision | When |
| - | - |
| **Sustain** | Nothing material changed. Keep monitoring, and keep ownership active. |
| **Update** | The work changed. After an approved update, save the new baseline, run the evaluations and record the release result. |
| **Build the next one** | The agent is stable and the next opportunity on the roadmap is ready. |
| **Re-agree the objective** | The outcome itself changed. Go back to Define. |
| **Scale** | It works where it runs. Qualify the next team or local variant of the process before extending it. |
| **Retire** | It no longer fits the work. Reconcile open cases, restore the fallback and close its access. |

Keep the previous version available, so you can roll back under your deployment controls.

## Who does what

| Role | Their part |
| - | - |
| **AI Builder & Deployer** | Maintains the agent, assesses drift and qualifies changes |
| **Functional Champion** | Refreshes the context and prioritizes improvements |
| **Executive Sponsor** | Decides to sustain, update, scale, build again or retire |
| **Process Performers** | Flag changes in the work and onboard their peers |
| **Within** | Reviews the value and supports the next cycle |

## Where to go next

> **[Measure the value](/user-docs/deliver/measure-the-value)**
>
> Show the agent is still paying off.

> **[Keep agents current (developer docs)](https://developers.within.ai/transformation/maintain-agents)**
>
> The developer view, with the MCP tools behind each step.

> [!NOTE]
>
> **Need help?** Use the in-app chat — click the chat bubble in the bottom-right corner of Within (staffed 24×5) — or email [support@within.ai](mailto:support@within.ai).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.