← All case studiesCase study 02 / Snorkel AI benchmarks

Writing the failure modes agents are tested against.

Task authoring and review for Terminal-Bench and Senior SWE-Bench. Reproducible environments. Independent verification. Real engineering failure modes.

Role
Expert Contributor
Organization
Snorkel AI
Scope / period
Software engineering & frontier AI
Explore benchmark projects

The context

A coding agent can produce a plausible patch without fixing the underlying problem. Evaluating it requires a reproducible environment and acceptance criteria that distinguish a working solution from a convincing response.

My benchmark work spans Terminal-Bench 2.0, 2.1 and 3.0 and Senior SWE-Bench. My contribution is task and review work within those wider collaborations.

Evaluation flow: task environment branches to reference solution and agent attempt. Both are checked by an independent verifier.Task + DockerReferenceAgent attemptVerifier
Evaluation structure showing separate reference and agent execution paths. An independent verifier assesses the resulting behavior.

My contribution

Authored tasks and proposals. Built containerized environments and engineered reference solutions. Hardened separate verifiers through behavioral validation and edge-case testing.

Worked on failure scenarios covering cluster hardening, insecure rollouts, RAFT state recovery, distributed tracing and infrastructure migration.

Reviewed repository-scale work across Better Auth, Apache Superset, Prefect, PostHog, SpacetimeDB and Gitea. Contributed reviewer guidelines and grading calibration.

Associated with Snorkel AI

Benchmark projects

These projects connect my task-authoring and review work to different questions about coding-agent reliability. Explore each benchmark for its methodology and published results.

Senior-level software engineering

Senior SWE-Bench

Senior engineering work drawn from real pull requests across 12 open-source projects. Agents investigate runtime failures, build features from realistic instructions and produce code that fits the surrounding repository. Evaluation considers functional correctness and engineering judgment.

Containerized terminal tasks

Terminal-Bench 2.0

A set of 89 curated tasks in isolated Docker environments. Each task includes a human-written reference solution and tests that verify the final container state. The benchmark originated as a Stanford University and Laude Institute collaboration with community contributors including Snorkel AI.

Benchmark reliability & validation

Terminal-Bench 2.1

A reliability update to the 2.0 task set. It repairs 28 of 89 tasks by addressing changed external dependencies, execution budgets and mismatches between instructions and tests. Continuous validation helps keep the benchmark reproducible as its dependencies evolve.

Evolving tasks & adversarial review

Terminal-Bench 3.0

Formerly Frontier-Bench. A harder successor to 2.1 that covers a broader range of work through open community contributions. Tasks undergo implementation review, reference-solution and no-op checks, agent runs and adversarial testing. The task set evolves as agent capabilities improve.

Engineering decisions

01. Test the evaluator itself

Oracle, no-op and partial-fix checks establish whether a verifier recognizes a correct solution and rejects an incomplete one.

02. Keep verification independent

Separating the verifier from the reference solution reduces answer leakage and keeps grading focused on required behavior.

03. Calibrate the review process

Clear reviewer guidelines help different reviewers apply the same acceptance criteria consistently.

The outcome

A body of task-authoring, environment-building and review work focused on reproducibility, correctness and regression risk. The engineering contribution extends to the checks used to judge the agent.

Tools & evidence

DockerPythonBashKubernetesAdversarial testingRegression analysisVerifier design

The linked project pages describe the wider benchmarks and their results. My contribution is the task-authoring, environment-building and review work described here.

NEXT CASE STUDYTreetracker by Greenstand
Let's start a conversation

What are you trying
to keep running?

Have a role or a project in mind?
Tell me what you’re working on.

North Carolina, USAOpen to remote opportunities

Send me a message

Minimum 10 characters.

Privacy Policy