Assistant instructions and settings

An assistant's saved instructions and settings are its agent definition: the maintained version of how that assistant does its job.

Northwind does not need a new copy of Scout for every person. An authorized builder can improve Scout's definition, and later work with Scout reads that saved version. Earlier conversations remain as they happened.

Write instructions someone could follow

Write the standing instructions as if you were briefing a capable colleague on their first day. They need to know the job, the expected result, and where not to guess.

For Scout, Northwind might save:

Prepare a daily operations brief from facts in the request. Start with Today.
Include Schedule and Risks. Do not invent missing owners, causes, dates, or
work. If a section has nothing to report, write None reported. Keep the brief
under 120 words.

This leaves the day's visits and delays in the current request, where they can change. It keeps the stable rules on Scout, where every intended user gets the same version.

Put a procedure used by several assistants in a skill. Keep changing business data in a maintained table, file, or source system. Use shared memory for knowledge someone deliberately promotes beyond one conversation.

Pick a model after the task is clear

The model is the AI itself. Start with the available default and get the test case right before comparing another model. Change one variable, send the same facts again, and inspect the result and cost.

The saved model is Scout's default. A new conversation uses it unless the person starting that conversation deliberately chooses another available model. That explicit choice creates a model_override for the session and keeps that conversation on the selected model.

If a session has no override, future turns inherit Scout's current saved model. Changing Scout's model can therefore affect the next turn in an existing unpinned conversation. A session with a model_override stays on its selected model. Worker-routed agents do not accept session model overrides.

Changing the model does not copy or replace Scout's instructions, tools, skills, memory policy, or timeout. Retest any tool use and structured output that depends on model behavior.

Know what an edit changes

A definition includes the assistant's instructions, model and reasoning setting, turn timeout, tools, skills, memory policy, and related configuration. Its audience and account grants are attached controls with their own authority checks.

After an authorized person saves an edit:

• future turns use the current saved instructions and settings;
• an earlier request and reply remain unchanged in their session;
• the assistant keeps the same ID, conversations, and recorded history; and
• a different assistant is unaffected.

This makes one shared definition maintainable without pretending that history was always produced by the latest version.

Improve one thing at a time

1. Start a fresh conversation with the assistant and send a small task with known facts.
2. Check the facts, format, tool calls, and final status.
3. Open Build > Agents, choose the assistant, then click Edit on Configure.
4. Change one instruction, model choice, tool, or limit.
5. Click Save.
6. Start another fresh conversation and send the same task.

The second run should show the intended change without introducing a new unsupported claim. If it does not, the two conversations give you concrete input and output to compare.

Editing requires more than audience membership. A colleague may be allowed to use Scout without permission to read its private configuration or change the definition. Keep maintenance with the people who own the test cases and access decisions. See who can use an assistant.

Create another assistant when the boundary changes

Edit Scout when you are improving the same job for the same group. Create a separate assistant when the work needs a different audience, different accounts, or knowledge that must stay apart. A new name alone is not a useful boundary; the people and access behind it are.

For example, Northwind's operations brief and an IT provider's client-ticket triage may both summarize text. They should still be separate assistants if they use different customer data, tools, and reviewers.

Keep one definition authority

Mobius stores the agent definition. Conversations, routine occurrences, messaging channels, and API invocations read that stored configuration. An invocation supplies the current request; it does not send a second inline definition.

Software that provisions assistants can use the assign-once external_ref to find the same agent again and update its saved configuration deliberately. Exact fields, model override rules, and request examples are in the Agent configuration API guide.