Agent Design

How Stella's agent actually behaves: system prompts, tool definitions, context management, and how we measure whether changes make the agent better or worse.

🚧 Starter page — this is the pillar for all prompt/behavior engineering knowledge.

System prompt

Where the prompt lives in code, how it's versioned, and the process for changing it (prompt changes should always run through evals before merge).

Tools

Tool	Purpose	Definition
<tool name>	<what it does>	<link to code>

Guidelines for adding a tool: keep descriptions action-oriented, define clear input schemas, and add at least one eval case per tool.

Context management

How we build the context window: what gets included, truncation strategy, memory/summarization approach.

Evals

How to run the eval suite, where results are tracked, and the bar a change must clear. Include known-flaky cases and how to interpret variance.

Prompt-engineering lessons learned

Append dated entries — this section is the team's collective memory:

• YYYY-MM-DD — <what we tried, what happened, what we do now>

Related pages

• Architecture — where the agent loop sits in the system
• Decision Records — rationale behind major behavior choices