| --- |
| license: cc-by-4.0 |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - devops |
| - sre |
| - software-engineering |
| - benchmark |
| - agents |
| - mcp |
| - deterministic-evaluation |
| pretty_name: DevOpsBench-100 |
| size_categories: |
| - n<1K |
| --- |
| |
| # DevOpsBench-100 |
|
|
| DevOpsBench-100 is a synthetic long-horizon software-engineering / SRE agent |
| benchmark: 100 tasks over one executable world ("NovaCart", a mid-size |
| e-commerce SaaS) with 72 SQLite tables, |
| 1451 seeded rows, a 38-file monorepo with 417 commits, |
| and 97 MCP tools spanning a first-party engineering stack |
| (tickets, PRs, CI, deployments, canaries, migrations, feature flags, metrics, |
| alerts, incidents, chat, knowledge base) plus deliberately disagreeing |
| vendor-shaped surfaces (Jira, Linear, GitHub Issues, Prometheus, Sentry, |
| PagerDuty, Confluence, spreadsheets) and Kubernetes. |
|
|
| Tasks are high-level workplace requests: the employee states the operational |
| question or desired business result while procedure, authority, and causal |
| facts remain distributed across the sandbox. Reference trajectories run |
| 32-67 tool calls (median 35), with |
| 100/100 distinct tool-name sequences. Every task publishes 53 |
| task-scoped native assets: 32 contextual workplace artifacts, 18 exact |
| non-empty sandbox exports that correspond one-to-one with the |
| decision-controlling reads, and one inspectable asset manifest. The reference performs |
| 26-34 |
| context reads, while the verifier requires all 20 |
| decision-controlling joins, the task-specific state transitions, a derived |
| capacity answer with three date/cost/authority options, provider readback, and |
| a reopened handoff. Acceptance is fully deterministic and expressed as |
| 16 task-specific semantic milestones totaling 100 points. |
| Low-level vcode, final-state, sequence, and anti-forgery checks remain nested |
| verifier evidence - no LLM judge, network, or clock is in the reward path. |
|
|
| ## What is included |
|
|
| - `data/tasks.jsonl`: task records (`task_id`, `task_name`, `world_id`, |
| `prompt`, `context_files`, `rubric`, `gold_output`, `metadata`). |
| - `tasks/`: one readable JSON record per task (includes the guided |
| instruction variant). |
| - `task_files/`: 53 inspectable assets per task spanning tickets, services, |
| observability, incidents, deployments, code, CI, migrations, vendor trackers, |
| knowledge, chat, approvals, capacity reservations, vendor lead times, and |
| customer commitments. The `material/` subdirectory contains the |
| 20 |
| exact sandbox exports used by the deterministic causal contract; the manifest |
| records source, query scope, material reason, format, byte size, and digest |
| for every listed file without disclosing an execution order or gold state. |
| - `world/`: the offline world source - stdlib MCP server, tool implementations, |
| seeded SQLite database, schema and seed SQL. |
| - `verifiers/`: 100 standalone verifier scripts |
| (`python3 verify_<task>.py world.db` prints the full verdict). |
| - `trajectories/`: one executed reference trajectory per task (JSONL). |
| - `reports/`: measured build and qualification evidence. |
|
|
| ## Task families (19) |
|
|
| | Family | Tasks | |
| |---|---:| |
| | AIOps root-cause analysis | 17 | |
| | Cross-system source of truth | 13 | |
| | Error-rate SLO recovery | 7 | |
| | Feature-flag operation | 7 | |
| | Multi-service rollout | 7 | |
| | Cross-source reconciliation | 7 | |
| | Latency optimization | 6 | |
| | Security incident response | 6 | |
| | AIOps localization | 5 | |
| | Flaky-test remediation | 5 | |
| | AIOps detection | 4 | |
| | Code implementation | 4 | |
| | Operational judgement / restraint | 4 | |
| | Incident handover | 2 | |
| | Long-horizon delivery | 2 | |
| | API migration | 1 | |
| | Change attribution | 1 | |
| | Human-approval gated change | 1 | |
| | Workspace scripting | 1 | |
|
|
| ## Objective release gates |
|
|
| | Gate | Required | Measured | |
| |---|---:|---:| |
| | Tasks | 100 | 100 | |
| | High-level unique employee requests | 100 | 100 | |
| | Unique reference tool sequences | 100 | 100 | |
| | Inspectable assets per task | 53 | 53 | |
| | Exact non-empty material exports per task | 20 | 20 | |
| | Material causal reads per task | 20 | 20 | |
| | Unique causal evidence profiles | 100 | 100 | |
| | Semantic milestones / points | 16 / 100 | 16 / 100 | |
| | Oracle replays at reward 1.0 | 100/100 | see `reports/qualification.json` | |
| | Deterministic verifier replays | 100/100 | see `reports/qualification.json` | |
| | Negative-control false accepts | 0 | see `reports/qualification.json` | |
| | LLM / network calls in verifier | 0 | 0 | |
|
|
| ## Provenance and contamination |
|
|
| Every service, metric, document, commit, and incident is synthetic and was |
| generated for the NovaCart world. 30 cross_system/handover tasks were ported |
| from TheAgentCompany task *shapes* re-grounded onto NovaCart's own state; the |
| AIOps families reproduce the microsoft/AIOpsLab task *structure* |
| (detect / localize / analyze) against NovaCart. No third-party benchmark |
| text, evidence, or gold answers are included. Gold outputs are public, so |
| this release is appropriate for transparent evaluation and RL experiments |
| rather than secret-test claims. |
| |
| ## Licenses |
| |
| Synthetic task data and world content are CC-BY-4.0. Benchmark code, |
| server, and verifiers are Apache-2.0. |
| |
