Working with the agent
How to work effectively with the Sazabi agent — the artifacts it produces, steering long investigations, prompting with project memory in mind, and referencing your project's objects.
The Sazabi agent does the querying and correlating; your job is to point it at the right problem and read what it produces. This page covers the artifacts it renders, how to steer a long investigation, how project memory shapes what you should ask, and how to reference the objects in your project.
Artifacts the agent produces
The agent does not only reply in text. When it helps, it renders artifacts inline in the thread:
- Charts. Time-series visualizations of what it found — for example error rate over the last hour.
- Tables. Structured results, such as the top failing endpoints with counts.
- Code. Snippets and diffs, including references to the exact file and line when your code is connected.
- Pages. Longer formatted write-ups the agent assembles to explain what happened.
Every artifact is grounded in the queries the agent ran, and you can expand the tool calls to see the underlying data.
Steering long investigations
A good investigation is a conversation, not a single prompt.
- Start with the symptom and a time window ("checkout errors are up since 14:00 UTC — what changed?").
- Read the agent's tool calls as it works; if it is looking in the wrong place, say so and it will adjust.
- Ask follow-up questions in the same thread so the agent keeps the context it has built.
- If a thread has gone down the wrong path, fork it from an earlier message and take a different angle.
How project memory shapes what you ask
The agent draws on project memory — durable notes and runbooks your team has curated. When memory holds the context (which service owns checkout, how you define "healthy"), you can ask higher-level questions and skip the setup. When it does not, give the agent that context in your prompt, and consider writing it to memory so the next investigation starts ahead.
Referencing issues, components, and log streams
Refer to your project's objects by name or ID in your prompt, and the agent looks them up as part of its investigation — for example "investigate issue about failed webhook deliveries" or "show me logs from the checkout service." The agent resolves these against your project's data rather than requiring a special syntax.