---
title: "AI engineering · Ataris"
description: "Ataris records the tool calls an agent made, what each one returned, and what each command did on the machine, then ties every action to the decision that caused it."
canonical: "https://ataris.dev/solutions/ai-engineering"
last-updated: 2026-10-10
---
# See the context behind every decision your agent made.

Ataris records the tool calls an agent made, what each one returned, and what each command did on the machine, then ties every action to the decision that caused it.

## Debugging a bad decision

1. **Open the run.** The session's mission sits at the top of its timeline: the one you passed, or one inferred from the agent's first prompt.
2. **Read what the model saw.** The context lane shows content entering the model's context, such as a README the agent read or a tool result it got back.
3. **Follow the tool call.** Each tool call sits next to its result, and tool-use links connect it to the processes it started and the files and hosts they touched.

## What AI engineers get

- **Decisions on the timeline.** Each decision to call a tool is recorded on the model lane, with the prompt that started the run at the top.
- **Context you can inspect.** Content entering the model's context is recorded as its own event, so you can see what the agent read before it decided.
- **Harnesses you already use.** Put ataris in front of Claude Code, Codex, or Grok to trace it where it already runs: a terminal, a CI job, or your own sandbox.
- **Runs you can compare.** Sessions are tagged with the agent's name, so you can line up runs of the same agent across prompts and versions.

## Trace Claude Code where it runs

Put ataris in front of the command. Each command, file, web fetch, skill, and MCP call the agent makes lands on the timeline.

```bash
export ATARIS_API_KEY=atk_...

ataris --agent test-fixer claude -p "Fix the failing tests"
```

## Common questions

### Are model calls recorded?

The CLI records the prompt that started the run, each decision to call a tool, and each result, but not the model's own text. Its calls to the model provider still appear as network events.

### Do I need to set a mission?

No. Pass one if you want, or Ataris infers it from the agent's first prompt.

## Debug the decision behind the run.

Trace your agent with the CLI, or book a demo to see a run replayed.

[Book a demo](https://ataris.dev/contact) · [Read the docs](https://ataris.dev/docs/guides/cli.md)