An autonomous research lab for agentic systems.

If you can run it, Bunsen can eval it.

Bunsen is a universal eval lab you can point at any task: benchmarks, zero-to-one product builds, coding agents, your own Claude Code customizations, and more. In fact, we've ported Terminal Bench 1.0 to Bunsen — and, as an example, we had Claude Code use the bn CLI to create BattlesnakeBench, where models compete to create the best Battlesnake bot.

Games · BattlesnakeBench

Seriously — eval anything.

Claude Code used the bn CLI to create BattlesnakeBench — here are two models' bots, each scored by win-rate^* against a hidden, held-out ladder of reference snakes:

^* Win-rate over 36 games per model (9 seeds × 4 hidden reference bots). The official engine spawns snakes nondeterministically, so a single run varies by ~±3% (1 SD) — the model ranking holds across repeated scorings. Each replay is one game, not the full ladder.

Maximum power, minimum setup.

Install the CLI, scaffold a project with the frontier coding agents already wired up, then write your first experiment — a folder with a task and a rubric. Every run is fully instrumented — traces, artifacts, diffs, and cost — captured automatically, with zero changes to the agent.

$schema: …/experiment.v1.json
version: v1
name: gcal-clone
task:
  prompt: Build a calendar app in /workspace — month / week / day views.
workspace:
  sources:
    - path: ./workspace
evaluation:
  criteria:
    - id: builds
      type: script
      run: bash /bunsen/verifiers/build.sh
      scores: [0, 1]
    - id: views
      type: browser-agent
      instructions: Make sure the month / week / day views work.

Deep by default.

Agentic scorers — agents that investigate and evaluate the artifacts produced by a task — are the heart of Bunsen's eval system. Bunsen grades the agent by giving it a real task and digging into what it produced. Scorers run a spectrum: from deterministic checks to AI agents that explore the workspace, drive a browser, and read the run's own traces, each citing its evidence. Because agentic scorers are expensive, Bunsen lets you gate them behind $0script scorers — paying for deep judgment only where it's warranted.

evaluation:
  criteria:
    - id: builds         # gate — skip the rest unless it compiles
      type: script
      run: bash /bunsen/verifiers/build.sh
      scores: [0, 1]       # allowed values — omit for continuous 0–1
      gate: { ifBelow: 1 }
    - id: security       # an agent audits the source and cites file:line
      type: agent
      instructions: Audit for injection / XSS — unsanitized titles, an unsafe ICS export.
    - id: drag-and-drop  # a browser-agent drives real Chromium
      type: browser-agent
      instructions: Drag to create an event, move it to another day — does it stick?
    - id: responsive     # graded 0 / .25 / .5 / .75 / 1, not pass/fail
      type: browser-agent
      scores: [0, 0.25, 0.5, 0.75, 1]
      instructions: Does the layout hold at 375 / 768 / 1280 px?
    - id: overall        # pure-math roll-up, no LLM
      type: aggregate
      needs: all
      aggregate: { function: weighted_average }

Cheap checks gate the expensive ones — a deterministic script short-circuits a browser-agent, so you only spend judgment where it matters. Skipped is recorded as skipped, never a fake zero.

Toward an autonomous research lab.

Bunsen runs on its own agents: inside every run, they drive the agent under test, read its traces, and score the result. The bigger goal is an autonomous lab that runs the whole research loop — proposing the questions, running the matrix, and writing up the findings. We're not there yet, but the last two steps already work: running the matrix produced the sweep below, and turning those runs into a cited write-up is what Bunsen did next.

Upcoming — a productized meta-analysis command (bn report) and a fully self-directed loop where Bunsen proposes its own research questions.

From agent to evidence.