AgentLens

See what AI coding agents really do — and learn when to trust them.

AI agents plan, edit files, run commands, and report "done" on their own — but a lot happens under the hood. AgentLens is a VS Code extension (and no-install browser app) that lets you click through what an agent actually did, step by step: open each step to reveal the hidden details, see the safety flags, and decide — approve, replan, or block. It's built to be quick, visual, and friendly for CS and non-CS students alike.

No setup, no real commands, nothing to break. Sample runs are safe simulations; you can also import a real run from your own repo.

Two ways in

• 🔍 Trace Explorer — the heart of it. A run is shown as a summary chain of step nodes (green / amber / red by risk) plus a detailed thread. Open any step to see what really happened, with an oracle-computed safety verdict, then make the call. Explore the bundled sample runs, or import a real SWE-agent run (.traj).
• ⚡ Trust scenarios — short, gamified what-would-you-do moments (the Phantom Revert, the Overeager Cleanup, …) that score your trust calibration: are you over-trusting, or over-cautious?

Getting started

1. Install the extension.
2. Click the lens icon in the Activity Bar → Open AgentLens, or run AgentLens: Open from the Command Palette (Ctrl/Cmd+Shift+P).
3. Explore a run or play a scenario. Your trust-calibration profile builds as you go.

Bring a real run

You can explore a run an agent actually performed on a real repository. AgentLens reads SWE-agent trajectory files (.traj / .json / .jsonl): each step's thought becomes the intent, the action the command, the observation the under-the-hood detail, and the reference oracles flag anything risky.

There are two commands — one to import a run you already have, and one to produce a run.

A. Already have a .traj? Import it

1. Command Palette → AgentLens: Import a Real Agent Run (SWE-agent)….
2. Pick the .traj (or .json/.jsonl) file. SWE-agent writes it under your output directory (see below) — e.g. swe-agent-output/<run-id>/<run-id>.traj.
3. The run opens in the Trace Explorer — chain view + thread, with safety verdicts.

B. Run SWE-agent on this repository

This produces a real trajectory you can then import. AgentLens does not run the agent for you — it builds the exact command and drops it (un-executed) into a terminal so you can read it before running. SWE-agent runs in its own sandbox; AgentLens only reads the result.

Prerequisites (one-time):

# 1. Install SWE-agent (see https://swe-agent.com for the current instructions)
git clone https://github.com/SWE-agent/SWE-agent.git
cd SWE-agent && pip install --editable .

# 2. Give it an API key for the model you want it to drive
export ANTHROPIC_API_KEY=sk-ant-...      # or OPENAI_API_KEY, etc.

Each run:

1. Open the repository you want the agent to work on as your VS Code workspace folder.

2. Command Palette → AgentLens: Run SWE-agent on This Repository….

3. Type the task (the problem statement), e.g. "Fix the failing test in the auth module."

4. AgentLens opens a SWE-agent terminal pre-filled with a command like:

   sweagent run \
     --agent.model.name claude-sonnet-4-5 \
     --env.repo.path "/path/to/your/repo" \
     --problem_statement.text "Fix the failing test in the auth module" \
     --output_dir swe-agent-output
   

   It is not auto-run — review it first. Change --agent.model.name to your preferred model (e.g. another Claude or an OpenAI model your key supports). Press Enter to start it.

5. When it finishes, SWE-agent writes a .traj under swe-agent-output/…. Run AgentLens: Import a Real Agent Run… and pick that file to explore what the agent did.

Tip: agents can run shell commands and edit files in their sandbox. Point SWE-agent at a clone or a disposable checkout, not anything precious — and treat every step in the trajectory as something to review, which is exactly the skill AgentLens is teaching.

Privacy & safety

Runs 100% locally. No telemetry, no network calls. Sample scenarios and traces are scripted simulations — the agent never touches your files or shell. Imported runs are read-only.

Build from source

npm install            # from the monorepo root (workspaces)
npm run compile -w agentlens
npm test -w agentlens
npm run preview -w agentlens   # no-install browser build (great for demos / STEM day)
npm run build:vsix -w agentlens

Where it fits

AgentLens is the trace-exploration & trust tool. Its sibling Agentic-SE Coach is the assessment tool for semester-long, group agentic-SE projects (it scores a real repo on the Supervision Skill Model). Both are built on the shared @agentsafe/core library.

License

MIT © Sumon Biswas.