Playground Explorer

Mission

Explore how the AI agent behaves on a subject through iterative questioning, inspect trajectory artifacts as they are produced, and propose environment improvements that make better trajectories easier.

The playground is the environment in the observation-action loop where tools and UI are the agent interface:

observe → decide → act → observe ...

Primary execution path (CLI-first)

Use the playground CLI as the default path for all probes:

uv run python -m varro.playground.cli

Optional:

uv run python -m varro.playground.cli --user-id 1 --chat-id 62 --current-url /dashboard/boligpriser-sammenligning

Use commands during exploration:

• •:status
• •:url <path>
• •:trajectory [turn_idx]
• •:snapshot [url]

Avoid non-CLI trajectory generation unless blocked.

Workflow

1. •Start or resume a CLI session.
2. •Define a subject and exploration goal.
3. •Ask one probe question at a time, like a human investigator.
4. •After each turn, inspect:
   • •final response,
   • •trajectory artifacts (turn.md, tool_calls/, images/),
   • •URL/snapshot artifacts when relevant.
5. •Record what changed in behavior and why.
6. •Continue targeted follow-up probes until root causes are clear.
7. •Write or update findings.md.

Exploration rubric

Do not start from "was the answer correct?" Start from trajectory mechanics:

• •What observation did the agent receive?
• •Why did that observation lead to this action?
• •What missing or ambiguous signal created extra steps or uncertainty?
• •What is the smallest change in tool output, prompt, docs, or UI that would improve the next decision?

Correctness matters when it reveals environment gaps.

Scope boundaries

In scope:

• •Tool output shape and clarity
• •Tool instructions and prompt clarity
• •Documentation discoverability and task-fit guidance
• •URL/state ergonomics
• •Snapshot usability
• •Dashboard as communication surface for agent and user

Out of scope:

• •Broad model capability judgments without environment evidence

Output contract

Always write (or update):

data/trajectory/{user_id}/{chat_id}/findings.md

Each finding must include:

• •Hypothesis
• •Probe (question + optional URL context)
• •Observed Trajectory Evidence (turn/step/tool refs)
• •Interpretation
• •Proposed Change
• •Expected Trajectory Delta
• •Validation Probe

Template:

# Findings: Chat {chat_id}

## Exploration Goal
{What subject/behavior was explored and why}

## Findings

### {short title}
**Hypothesis**: {what you expected}
**Probe**: {question asked, optional URL}
**Observed Trajectory Evidence**: {turn/step/tool references}
**Interpretation**: {why behavior emerged}
**Proposed Change**: {specific system change}
**Expected Trajectory Delta**: {steps removed / decision clarified}
**Validation Probe**: {next question to verify change}

## Prioritized Actions
1. {highest impact, lowest effort first}
2. {next}

Stopping criteria

Stop when one of the following is true:

• •Root causes are clear and supported by trajectory evidence.
• •Additional probes are not changing conclusions.
• •You have at least one high-impact, testable proposed change with a validation probe.

Relationship to other skill

Use $analyse-trajectory for retrospective audit of completed chats. Use $playground-explorer for interactive probing and improvement discovery.