Skip to content

Latest commit

 

History

History
156 lines (119 loc) · 7.99 KB

File metadata and controls

156 lines (119 loc) · 7.99 KB

Why ToolTrace Bench — and when to use something else

A positioning document, written to be useful to somebody deciding rather than to somebody already sold. It ends with the cases where a different tool is the right answer, because a comparison that never reaches that section is marketing.


The one-sentence version

Others score the tool call, or watch production, or probe security. ToolTrace Bench scores the whole run — correctness, recovery, side effects, cost, latency and security resilience — and hands you a checksummed artifact a third party can reproduce and an auditor can accept.

The problem this exists for

Half of enterprises surveyed in June 2026 had shipped an agent that passed internal evaluation and still caused a customer-facing failure; one in four more than once. Gartner expects 40% of enterprise AI failures by 2028 to trace to inadequate evaluation and monitoring rather than to model capability.

The gap is not that agents are unmeasured. It is what gets measured. A benchmark that reports one accuracy number cannot distinguish:

  • an agent that solved the task from one that deleted the test file;
  • an agent that recovered from a tool failure from one that never hit one;
  • an agent that costs $0.04 a resolved task from one that costs $4.00;
  • an agent that ignored a prompt injection from one that was never shown one;
  • a real 4-point improvement from noise on twelve runs.

Those are different questions, and each needs a different measurement taken during the same run. Taking them separately, later, on different samples, is how you get a green dashboard and a customer-facing failure.

What "scores the whole run" means concretely

Every tooltrace run produces a .tooltrace bundle: six files plus a manifest of SHA-256 checksums, holding the task exactly as it was executed, the full trace, the final workspace diff, the score with every assertion, and the environment it ran in.

From that one artifact:

Axis What is measured Command
Correctness Deterministic assertions against the final workspace tooltrace run
Trajectory Tool-call AST matching, read-before-write ordering, loop detection, hallucinated resources tooltrace trace
Recovery Graded: recovered immediately, recovered slowly, or silently wrong tooltrace perturb
Side effects Blast radius — what the agent touched versus what it could have tooltrace verify
Cost Cost per resolved task, with unpriced runs excluded rather than counted as free tooltrace cost
Latency Inference time and tool time split, with hardware recorded tooltrace hardware
Security Attack success rate per class, with Wilson intervals tooltrace owasp
Evidence Dated, hash-chained dossier mapped to regulatory obligations tooltrace evidence

Where the categories actually divide

Twenty-six projects are tracked in competitive-analysis.md, fetched from the GitHub API and committed to data/competitor-meta.json so the table cannot drift from the data it cites. They fall into three groups that are easy to confuse:

Task suites define problems. SWE-bench, tau-bench, OSWorld, WebArena, BFCL. They are not competitors — importers ship for four of them (tooltrace import --format swe-bench|bfcl|tau-bench|agentbench), because a problem set and a harness are different things and you need both.

Eval harnesses run and grade. inspect_ai, promptfoo, DeepEval. This is the category ToolTrace Bench is in, and the difference is what it grades on: the execution trace and the final workspace, never a model's opinion of its own work.

Tracing platforms record production. Langfuse, Phoenix, AgentOps, Datadog. Integration targets, explicitly — tooltrace platforms --target … exports to all of them. Be the harness that feeds them, not a worse version of them.

The four things nothing else does together

  1. Cost on the same axis as accuracy. Researchers note costs are rarely reported at all. Cost per resolved task is a headline metric here, and an agent whose runs were never priced is excluded from the frontier rather than plotted at zero — treating unpriced as free puts it on the Pareto frontier by default.

  2. Security resilience as a benchmark axis. Prompt injection is OWASP's #1 for agentic applications and reported attack success rates reach 84%. Exfiltration is measured through an offline egress sink: the http_post tool performs zero network I/O and appends to a log inside the workspace, so an exfiltration attempt becomes an ordinary workspace observation that ordinary scorers can check.

  3. Statistics that refuse to over-claim. showdown will not name a winner the sample cannot support. Comparisons use Newcombe difference-of-Wilson intervals; power reports the minimum detectable effect before you spend the money; variance decomposition separates the model's nondeterminism from the harness's.

  4. Evidence a reviewer can accept. Checksummed, tamper-evident, reproducible, hash-chained. The EU AI Act's operative demand is that organisations demonstrate compliance rather than assert it, and a bundle somebody else can re-run is the difference.

When to use something else

  • You need a red-teaming platform with a large prompt corpus and a UI for iterating on prompts. Use promptfoo. This project measures whether an agent's behaviour holds up, not whether a prompt phrasing is better.

  • You need a general evaluation framework to build your own scorers and model-graded rubrics on. Use inspect_ai. It is a framework; this is a reliability benchmark with opinions, and those opinions include refusing model-graded scoring.

  • You need to watch live production traffic with dashboards and alerting. Use Langfuse, Phoenix or Datadog, and export to them from here.

  • You want a model leaderboard for isolated function-calling accuracy. Use BFCL. It does that well and at a scale this project does not attempt.

  • Your agent is a pure text generator with no tools and no side effects. Most of what this measures does not apply. Use a conventional eval harness.

What this project will not do

Stated because a positioning document that only lists strengths is an advertisement:

  • No model-graded scoring. No LLM judges the output. Deterministic assertions only, which means some tasks cannot be expressed here at all.
  • No compliance determination. The evidence dossier assembles facts and names the obligations no benchmark can evidence. It never says "compliant".
  • No hosted service. Self-host it or run it locally. Your data stays yours because it never leaves.
  • No claim that the sandbox is a security boundary. It is a workspace boundary with documented limits; see threat-model.md.
  • No telemetry. Nothing phones home, and there is no account.

Verify all of this rather than believing it

Every capability claim in this repository is graded in feature-status.md — including E for "works, external validation blocked" and N for "does not exist" — and the table is machine-checked in CI. So are the README, the CLI reference, and every command in every documentation code block.

That is the actual differentiator, and it is the one worth checking first: if a project's documentation is not verified against its code, none of the rest of its claims can be either.

Further reading