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

# Agent Configuration

> How to configure any CLI tool as an Agent Bench agent

## Agent YAML Format

An agent config is a YAML file with two fields:

```yaml theme={null}
command: <shell command to run>
timeout: <ISO 8601 duration>
```

The command runs via `bash -c` in the workspace directory.
The agent should read `INSTRUCTION.md` and modify the workspace.

## Examples

### Claude Code

```yaml theme={null}
command: claude --print --dangerously-skip-permissions "Read INSTRUCTION.md and follow the instructions precisely."
timeout: PT45M
```

### Gemini CLI

```yaml theme={null}
command: gemini -p "Read INSTRUCTION.md and follow the instructions."
timeout: PT30M
```

### Shell Script

```yaml theme={null}
command: ./my-agent.sh
timeout: PT10M
```

Your script receives the workspace as its working directory:

```bash theme={null}
#!/bin/bash
# my-agent.sh
INSTRUCTION=$(cat INSTRUCTION.md)
# ... your agent logic here
echo "Hello World!" > hello.txt
```

### Python Agent

```yaml theme={null}
command: python3 /path/to/agent.py
timeout: PT15M
```

## The Filesystem Contract

When your agent runs, the workspace contains:

| File             | Description                                               |
| ---------------- | --------------------------------------------------------- |
| `INSTRUCTION.md` | The task description (always present)                     |
| Source files     | Workspace template files (if the benchmark provides them) |

Your agent should:

1. Read `INSTRUCTION.md` to understand the task
2. Create or modify files in the current directory
3. Exit when done (zero or non-zero exit code)

The benchmark grades the workspace contents after your agent exits.

## Optional: Agent Journal

If your agent writes a `journal.yaml` to the workspace, the benchmark parses it for efficiency metrics:

```yaml theme={null}
schema: bench.journal.v1
totalTurns: 8
totalInputTokens: 4000
totalOutputTokens: 2000
totalCostUsd: 0.12
durationMs: 15000
phases:
  - name: plan
    turns: 3
    inputTokens: 1500
    outputTokens: 800
    costUsd: 0.05
    durationMs: 6000
    toolUses:
      read: 5
      write: 2
```

Agents that don't produce a journal still get graded --- only efficiency metrics are missing.

## Optional: Trajectory Reference

If your agent writes a `trajectory-ref.txt` file containing a URI or path to trace data,
the benchmark records it in the trial result for later analysis.

```
s3://my-bucket/traces/run-123.jsonl
```