Skip to content

Try τ²-bench as an eval harness for drua #200

Description

@nicolasburtey

Idea

Adopt τ²-bench (Sierra) as our first external eval harness. It's the public benchmark whose shape lines up most cleanly with drua: a tool-using agent talks to a simulated user, calls real APIs against a stateful backend, and is graded by the resulting state diff (not LLM-as-judge alone).

Why it fits drua specifically

drua is fundamentally an MCP gateway routing an agent through real toolsets behind a policy prompt. τ²-bench's runtime shape is the same: tool catalog + policy + stateful backend + scripted user goal + deterministic grader. Most other public benchmarks miss our value prop:

  • SWE-bench / Terminal-Bench → pure code editing, doesn't exercise the gateway
  • WebArena / OSWorld → wrong surface (browser / desktop)
  • GAIA / AgentBench → mostly synthetic, not tool-orchestration shaped

τ²-bench gives us:

  • Pass^k reliability metric (not just best-case) — runs the same task k times, only passes if all k succeed
  • Deterministic state-diff grading — LLM-judge is only used for the user simulator, not the grade
  • Dual-control domain (telecom) — user also has tools, agent has to instruct rather than execute. Maps well to "ops engineer + drua co-debugging"

Proposed plan

  1. Baseline run — run airline + retail domains as-is through drua's gateway. Gives a public-comparable number and shakes out harness wiring. (~1-2 days)
  2. Custom drua domain — fork the task format with one of our toolsets. Best candidate: concourse (we already have benchmarks/concourse/follow-rollout as a seed task). Write ~20 scripted user goals like "find out why the deploy job failed and retry it." (~1 week)
  3. Compare to internal benches — we already have benchmarks/notification-domain (code_assistant ablation) and benchmarks/concourse. τ²-bench gives us the missing piece: a reusable, pass^k-graded harness instead of one-off scripts.

Out of scope for now

  • Voice / full-duplex evaluation (τ²-bench supports it; not relevant to us yet)
  • Multi-toolset tasks — start with one toolset per domain to keep grading deterministic

Links

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions