---

Why This Exists

The 2025 Stack Overflow Developer Survey shows 84% of professional developers using AI in their workflow, up from 76% the year before. Yet only 33% trust AI accuracy while 46% actively distrust it (survey.stackoverflow.co/2025/ai). METR's RCT of 16 experienced open-source maintainers found AI tooling increased task completion time by 19%, while developers self-reported a 20% speedup, a 39-point gap between perception and reality (arXiv:2507.09089). Static issue benchmarks like SWE-bench Verified measure model capability in isolation; SWE-bench Pro (arXiv:2509.16941) addresses contamination at scale but still scores patches, not whether a workflow can ship.

AWB measures whether a configured tool+workflow combination can ship correct, regression-safe, low-burden changes against pinned real-world repositories. The same model running vanilla Claude Code vs. a purpose-built setup with a tuned CLAUDE.md, hooks, and structured agents produces meaningfully different results on real engineering tasks. AWB benchmarks the full stack: tool + configuration + workflow + model, together, on 100 tasks drawn from real open-source repositories.

How AWB relates to other benchmarks

Related work measures complementary axes. HAL analyzes agent traces with LLM judges across 11 tasks at 2.5B-token scale. Artificial Analysis publishes harness-vs-harness comparisons holding the model constant. SWE-bench Verified and SWE-bench Pro score patches against real GitHub issues. LiveCodeBench addresses contamination by time-segmenting contest problems. METR RE-Bench compares humans and agents in matched ML-engineering environments.

AWB's distinct contribution is twofold: (1) a paired vanilla-vs-custom adapter pair that isolates the workflow-configuration delta for the same model, surfaced as a single Workflow Lift score with a sign-test p-value; (2) deterministic trace-grading rubrics (read-tests-before-edit, ran-verification-after-change, no-out-of-scope-edits, no-repeated-failing-loop) computed from OpenTelemetry-aligned .trace.jsonl artifacts, not LLM judges. See METHODOLOGY.md#related-work for citation details.

What's New in v1.4.0

• Trace grading actually grades now. The runner used to emit only token spans, so all four trace rubrics scored a vacuous 100 on every run. It now translates Claude Code's nested tool_use blocks into FILE_EDIT / read / SHELL_COMMAND spans and correlates Bash exit codes — producing real, discriminating scores (the published baseline's no_out_of_scope_edits ranges 17-100 across the 8 tasks).
• Baselines carry the trust columns: per-run trace_grade and a submission-level readiness + trace_summary block, so the published baseline showcases both flagship trust features. A span-less trace reports null, never a fake 100.
• -j N works on its own. -j>1 enables parallel mode (it used to be a silent no-op without --parallel); a crashed parallel task is now recorded as a FAIL with a traceback instead of vanishing from results.
• Aider is a real adapter (is_stub = False), gated on the binary being installed.
• Exact-pinned runtime dependencies for reproducible installs, plus invariant guard tests so the README lead, install pin, and baseline reference can't silently drift.
• Security posture documented in docs/SECURITY.md: the shell-execution trust boundary and the per-task Docker isolation planned for safe community submissions.

Carried over from v1.2.0-v1.3.0: public GitHub Pages leaderboard, task-set hash on every result, OpenTelemetry-aligned .trace.jsonl artifacts, awb trace grade, the Production Readiness Score, strict result schema v2, and reliability + provenance hardening.

Quick Start

pip install awb

awb quickstart                                        # verify your setup
awb warmup                                            # pre-build workspace templates (one-time, ~5 min)
awb run --fast-check claude-code-custom               # 8 tasks, ~15 min, ~\$4 (quick signal)
awb run --progressive --adaptive claude-code-custom   # full suite with early exit + smart re-runs
awb gap results/runs/<run_dir>/                       # analyze capability gaps
awb leaderboard --readiness --explain                 # Production Readiness Score per tool

Five-minute reproducible demo

Run this end-to-end against the published v1.4.0 fast-check baseline. Should finish in ~15 minutes for ~$4 of API spend and produce a tweetable Workflow Lift number plus a capability profile.

pip install awb==1.4.0
awb quickstart                                       # 1. verify environment
awb warmup --use-uv                                  # 2. pre-build templates
awb run --fast-check claude-code-custom              # 3. ~15 min, ~\$4, real run
awb leaderboard --readiness --explain                # 4. composite readiness score
awb trace grade results/runs/<run_id>/               # 5. behavior rubric scores

Compare against the published baseline at results/baselines/claude-code-custom-1.4.0-fast-check.json. Same task_set_hash means your numbers are directly comparable.

New in v1.1.0: awb warmup caches workspaces for 10-30x faster setup. --fast-check gives a quick signal in 15 min for ~$4. --progressive stops early on weak tools. --use-uv swaps pip for uv. See Execution Modes below.

How It Works

Clone repo at pinned SHA
  → Run setup commands
  → Capture baseline lint/security counts
  → Execute tool with task prompt
  → Run test suite + partial credit rubric
  → Sigmoid-normalize 7 metrics
  → Produce weighted composite + capability profile

Each task starts from a fresh git clone at a pinned commit. Every tool gets the same prompt, the same timeout, and the same verification suite. Results are scored with sigmoid normalization so scores are never negative and never collapse at the boundary.

Security: AWB clones third-party repos and runs their setup/test code plus the AI tool with no sandbox. Treat task sets and their repos as trusted input and run in a disposable environment. See docs/SECURITY.md for the trust boundary and the planned per-task Docker isolation.

Scoring System

Seven dimensions, sigmoid-normalized with per-task baselines derived from difficulty:

Dimension	Weight	What It Measures
Correctness	55%	Pass/fail (60%) + partial credit rubric (40%)
Cost efficiency	15%	Estimated USD per task
Speed	10%	Wall-clock seconds vs. estimated task time
Code quality	10%	Lint warning delta (pre vs. post)
Reliability	5%	Pre-existing tests broken by the change
Security	3%	New security issues introduced
Efficiency	2%	Blend of iteration count and tokens-per-iteration

Weight profiles (select with load_weight_profile(name)):

Profile	Focus	Use When
default	Balanced	Standard evaluation
correctness_focused	70% correctness	Research-grade rigor
production	45% correctness, 20% cost, 10% reliability, 8% security	Shipping to users
token_efficient	25% cost, 15% efficiency	Tight API budgets
rate_limited	30% cost, 15% efficiency	Hitting TPM/RPM limits

Sigmoid curve: score = 100 / (1 + exp(k * (value - baseline)))

• Optimal performance (excellent) → ~95
• Baseline performance (adequate) → ~50
• Above baseline → smooth decay, never negative

Difficulty-weighted aggregation: hard tasks count 2.5×, medium 1.5×, easy 1.0×. A tool that solves hard tasks beats one that only solves easy ones even if the easy-task count is higher.

Per-task baselines by difficulty:

Metric	Easy	Medium	Hard
Cost optimal / baseline	$0.05 / $0.30	$0.20 / $1.00	$1.00 / $3.00
Speed	50% / 100% of estimated_minutes	same	same
Iterations	3 / max_iters	8 / max_iters	15 / max_iters

The 100 Tasks

Real open-source repos, pinned to release tag SHAs. Setup runs in under 15 seconds via venv + pip.

Category	Count	Easy / Med / Hard	What It Tests
bug-fix	12	7 / 1 / 4	Root cause analysis, test-first diagnosis, N+1 queries
feature-addition	9	3 / 0 / 6	Convention adherence, ambiguous requirements, Dockerfiles, TypeScript typing
refactoring	11	5 / 2 / 4	Multi-file consistency, O(n^2) optimization, CI/CD config, async migration
code-review	9	4 / 2 / 3	Security review (report-only), concurrency analysis, migration guides, OWASP
debugging	10	7 / 0 / 3	Performance profiling, regression bisection, stack trace diagnosis
multi-file	7	4 / 0 / 3	Merge conflicts, plugin systems, auth chains
legacy-code	12	9 / 0 / 3	SQLAlchemy 2.0 migration, 20-file codebase navigation, dead code removal
workflow	30	9 / 12 / 9	Completeness tracking, convention discovery, security methodology, context utilization, async safety, config extraction, test-driven implementation

Repos used: FastAPI (74), httpx (17), Flask (4), Click (4), Starlette (1). All Python.

Task IDs: BF-001–014 · FA-001–010 · RF-001–012 · CR-001–010 · DB-001–011 · MF-001–009 · LC-001–012 · WF-001–030

Capability Profiles

Each task maps to 1–3 capabilities, producing a radar chart of tool strengths:

Capability	Tasks	What It Measures
code_comprehension	45	Understanding existing code before modifying
framework_knowledge	36	Knowing API patterns (Pydantic v2, async SQLAlchemy, etc.)
refactoring_discipline	29	Changing code without breaking behavior
bug_diagnosis	27	Structured root cause analysis, test-first diagnosis
multi_file_reasoning	22	Coordinating changes across multiple files
test_writing	12	Writing correct, meaningful tests
security_awareness	10	Identifying and fixing vulnerabilities
convention_adherence	8	Discovering and following project conventions
context_discovery	5	Reading project docs and config before editing
security_methodology	5	Applying security checklists systematically
completeness_tracking	4	Following all requirements, not stopping at 80%
cost_discipline	derived	Token efficiency across all tasks

Example awb gap output:

Capability Profile
------------------
code_comprehension    ████████████████████  82.4  (n=27, conf=high)
framework_knowledge   ████████████████░░░░  68.1  (n=26, conf=high)
refactoring_discipline████████████████░░░░  65.3  (n=23, conf=high)
multi_file_reasoning  ████████████░░░░░░░░  51.2  (n=20, conf=high)
bug_diagnosis         ███████████████░░░░░  63.7  (n=17, conf=med)
test_writing          ██████████░░░░░░░░░░  44.1  (n=8,  conf=low)
security_awareness    █████████████░░░░░░░  55.8  (n=8,  conf=low)

Systematic Patterns
-------------------
- Fails 70%+ of multi_file_reasoning tasks → consider multi-agent workflows
- Token spend on failed hard tasks: \$4.20 → add early-exit heuristics
- No failures on easy tasks → baseline is solid

Top Suggestions
---------------
1. Enable subagent mode for tasks spanning >3 files (impact: high)
2. Add repo-level CLAUDE.md with architecture overview (impact: medium)
3. Use --think flag for debugging tasks (impact: medium)

Vanilla vs Custom

AWB ships two Claude Code adapters that run the same model with different configurations:

	Vanilla	Custom
Hooks	Disabled	Your full hook suite
Skills	Disabled	Your registered skills
Auto-memory	Disabled	Active
System prompt	Generic	Default (loads CLAUDE.md)

Both use the same model, same API, same task prompts. The only difference is whether your workflow automation (hooks, skills, memory) is active. This isolates the contribution of workflow configuration from model capability.

Workflow Lift Score

When awb run executes both vanilla and custom (the default), it produces a Workflow Lift — a single number measuring how much your workflow configuration improves over the baseline:

Workflow Lift: +4.2 pts  (p=0.031, significant)
  Pass rate: vanilla 62% vs custom 68%
  Wins: custom 8 / vanilla 3 / ties 69

  Where your workflow helps:
    bug diagnosis             +12.3 pts  (17 tasks)
    multi file reasoning       +8.1 pts  (20 tasks)
    security awareness         +5.4 pts  (10 tasks)

  Where it hurts:
    cost discipline            -4.2 pts  (100 tasks)

  Biggest task-level differences:
    BF-014   +40  (V=35 C=75)
    LC-012   +15  (V=65 C=80)

The lift is computed per-task (configured score minus vanilla score), averaged across all tasks, and tested for statistical significance. Capability-level breakdowns show where your workflow configuration actually helps vs. adds overhead.

CLI Reference

awb run — Run benchmark tasks

awb run                            # all tools, all tasks, 3 runs (vanilla vs custom comparison)
awb run claude-code-custom         # single tool
awb run -t BF-001                  # single task
awb run --category legacy-code     # filter by category
awb run --difficulty hard          # filter by difficulty
awb run --capability bug_diagnosis # filter by capability
awb run --runs 1 --dry-run        # preview without executing
awb run --resume                   # skip tasks with existing results
awb run --parallel -j 4            # run 4 tasks concurrently
awb run --adaptive                 # re-run near-miss tasks (60-99%) after initial pass
awb run --progressive              # easy → medium → hard, stop early if pass rate too low
awb run --fast-check               # 8 representative tasks, 1 run (~15 min, ~\$4)
awb run --use-uv                   # use uv instead of pip for 10-30x faster installs

Execution Modes

AWB v1.1 ships four execution modes tuned for different evaluation scenarios:

Mode	Tasks run	Wall clock	Token cost	Use when
Full suite	300 (100 × 3 runs)	~3 hrs	~$150	Final evaluation, publishing results
Full + adaptive	~180	~1.5 hrs	~$100	Standard workflow, strong tools
Progressive	~150 on weak tools	~1 hr	~$40-75	Unknown/mediocre tools
Fast-check	8	~15 min	~$4	PR gates, iterating on config

Fast-check (8 representative tasks, 1 per category, reports estimated full-suite score ± margin):

Progressive (easy → medium → hard, stops if easy pass rate < 40% or medium < 20%):

--use-uv (rewrites pip install → uv pip install for 10-30x faster installs):

awb warmup — Pre-build workspace templates

awb warmup              # build templates for all 63 unique (repo, commit, setup) combos
awb warmup --dry-run    # show combos without building
awb warmup --clear      # reset template cache
awb warmup --use-uv     # use uv for faster initial builds

Workspace templates are cached at ~/.cache/awb/templates/. First build takes ~5 min; subsequent awb run invocations copy templates in ~2s instead of running pip install from scratch. Cuts ~55 min off a full benchmark run with 74 FastAPI tasks.

awb gap — Capability gap analysis

Analyzes results to produce a capability radar, failure classification, systematic patterns, and ranked improvement suggestions.

awb compare — Compare two runs

Side-by-side comparison of two benchmark runs with significance testing.

awb tools — List adapters

Shows all registered tool adapters and their availability status.

awb validate — Validate task YAMLs

Checks all 100 task YAML files against the schema, including partial credit sum-to-100 validation.

awb info — Task details

Displays full details for a specific task including repo, capabilities, and partial credit rubric.

awb stability — Score stability report

Per-task score variance across multiple runs. Flags unstable tasks for prompt clarification or tighter verification.

awb leaderboard — Generate HTML leaderboard

Generates a static HTML site with Chart.js radar chart, CSV export, and historical run tracking.

Add --readiness to print the Production Readiness Score per tool to stdout. The score is a weighted composite of correctness (35%), regression-safety (20%), security (15%), review-burden (10%), maintainability (8%), cost (7%), and speed (5%), all normalized 0-100. Weighted for shipping safety rather than headline accuracy.

awb trace grade — Score behaviors from trace artifacts

Every benchmark run writes a <task_id>_<tool>.trace.jsonl file using OpenTelemetry GenAI semantic conventions (gen_ai.client.operation, gen_ai.tool.use, gen_ai.usage.input_tokens) plus AWB-specific spans for shell commands (task.shell_command), file edits (task.file_edit), and test runs (task.test_run). awb trace grade <run_dir> reads each trace and scores four shipping disciplines on a 0-100 scale:

Behavior	What it checks
read_tests_before_edit	Did the tool read a test file before its first edit?
ran_verification_after_change	Was a test run / pytest invocation issued after the last file edit?
no_out_of_scope_edits	Did edits stay within files_to_examine from the task spec?
no_repeated_failing_command_loop	Did the tool retry the same failing shell command 2+ times?

awb calibrate-difficulty — Recalibrate difficulty labels

Recalibrates task difficulty labels from empirical pass rates. Use --apply to write changes back to task YAMLs.

awb calibrate-timeouts — Tighten timeouts

Recomputes task timeouts from empirical p95 wall-clock data. Use --apply to write changes.

Other commands

Command	Description	Demo
awb quickstart	Verify setup: tools available, tasks load	-
awb export <run_dir> -o file.json	Export results in submission format	-
awb submit <file.json>	Validate an external submission	-
awb compare-submissions <a> <b>	Cross-tool comparison with statistics	-
awb migrate-results <old_dir>	Convert v0.5.x results to v1.0 format	-
awb workflow <subcommand>	Export, validate, diff, or init descriptors	-
awb --version	Show version	-
awb run --dry-run	Preview tasks without executing	-

Adding Tasks

Tasks live in awb/tasks/<category>/. Copy awb/tasks/_template.yaml:

id: BF-012
category: bug-fix
title: "Fix response_model silently dropping extra fields in FastAPI"
difficulty: easy
estimated_minutes: 15
languages: [python]
capabilities: [framework_knowledge, test_writing]

repo:
  url: "https://github.com/tiangolo/fastapi"
  commit: "628c34e0"
  setup_commands:
    - "python3 -m venv .venv && source .venv/bin/activate && pip install -e '.[all]'"

issue:
  description: |
    The endpoint's response_model silently strips extra fields...
  files_to_examine:
    - "fastapi/routing.py"

verification:
  test_commands:
    - "source .venv/bin/activate && python3 -m pytest tests/test_extra_fields.py -v"
  partial_credit:
    - criterion: "Uses Pydantic v2 ConfigDict"
      points: 50
      check: "grep -q 'ConfigDict' tests/test_extra_fields.py"
    - criterion: "Tests pass"
      points: 50
      check: "source .venv/bin/activate && python3 -m pytest tests/test_extra_fields.py -v"

constraints:
  max_iterations: 20
  timeout_seconds: 1800

Run awb validate to check your task before opening a PR. Full guide: CONTRIBUTING.md

Supported Tools

Adapter	Name	Status
Claude Code (vanilla)	claude-code-vanilla	Full
Claude Code (custom)	claude-code-custom	Full
Pi	pi	Full
Gemini CLI	gemini-cli	Full
Codex CLI	codex-cli	Full
Cursor	cursor	Planned
Aider	aider	Planned
Windsurf	windsurf	Planned
Copilot	copilot	Planned

Run awb tools to see which are available in your environment.

Adding Tools

Implement the ToolAdapter ABC in awb/adapters/. v1.0 adds four optional methods to the ABC:

from awb.adapters.base import ToolAdapter, ToolResult
from pathlib import Path

class MyToolAdapter(ToolAdapter):
    name = "my-tool"
    display_name = "My Tool"

    async def execute(self, prompt: str, workspace: Path,
                      max_turns: int = 20, timeout_seconds: int = 1800,
                      on_event=None) -> ToolResult:
        ...  # on_event(event) callback for streaming token monitor; return False to abort

    def check_available(self) -> bool:
        ...

    def get_config_hash(self) -> str:
        ...

    # Optional — implement to enable pre-flight auth checks
    def supports_auth_check(self) -> bool: ...
    def check_auth(self) -> tuple[bool, str]: ...

    # Optional — implement to enable streaming metrics
    def supports_streaming(self) -> bool: ...
    def get_model_pricing(self) -> dict[str, float]: ...

Register in awb/adapters/registry.py and add an entry point in pyproject.toml.

External Submissions

Anyone can share results using the submission format defined in results/submission-schema.json:

awb run --runs 3
awb export results/runs/<run_dir>/ -o my-results.json
awb submit my-results.json                        # validate locally
awb compare-submissions a.json b.json             # compare with significance testing

The format captures tool version, model, hardware class, and per-task run results. Hardware classes (e.g., apple_m5_24gb, linux_x86_16gb) enable fair speed comparisons — only compared within the same tier.

Statistical Framework

• Confidence intervals via t-distribution (no scipy required for core scoring)
• Significance testing via sign test for paired tool comparison
• Integrity checks: contamination detection (completions <10s flagged), variance anomalies (identical times/tokens across runs)
• Weight profiles: default, correctness_focused, production, token_efficient, rate_limited (see awb/scoring/weights.yaml)
• Stability metric: per-task TaskStability (std_dev, score_range, is_unstable); high-variance tasks can be down-weighted in composite scoring
• Token efficiency: sigmoid normalizer (optimal=2k tokens/iter, baseline=15k) blended 50/50 with iteration count in the efficiency dimension

Changelog

1.1.0 (2026-04-07)

Performance and token optimization release. 33-50% faster full runs, ~97% cheaper quick evaluations.

• Workspace template cache — ~55 min saved on full runs (74 FastAPI tasks no longer re-run pip install)
• awb warmup — pre-build all unique workspace templates in parallel
• --use-uv — 10-30x faster pip installs via uv
• --progressive — easy → medium → hard execution, stops early if weak tool (50-80% token savings)
• --fast-check — 8 representative tasks, 1 run, ~15 min, ~$4 (97% cheaper than full suite)
• Token budget enforcement — max_input_tokens/max_output_tokens in task constraints, streaming kill switch
• Streaming token monitor — Claude Code adapter parses stream events as they arrive
• Parallel partial credit — independent grep/file checks run via asyncio.gather; pytest stays sequential
• Adaptive timeouts — runs 2+ tighten timeout to min(original, 2x run1_actual)
• Richer RunCost — cache_read, cache_creation, thinking token fields
• Token efficiency in scoring — efficiency dimension blends iterations + tokens-per-iteration
• Two new weight profiles — token_efficient and rate_limited for cost-sensitive evaluation
• Token-aware gap analysis — cost-per-point outliers, cache hit rate patterns, token burn detection
• JSONL results — additive output format alongside per-file JSON for fast batch loading
• 184 tests (up from 135)

1.0.9 (2026-04-04)

• Add Python 3.13 and 3.14 to CI test matrix and PyPI classifiers

1.0.8 (2026-04-04)

• Sync README changelog with PyPI long description; update GitHub repo description (80 → 100 tasks)

1.0.7 (2026-04-04)

Product audit fixes: 27 findings across observability, scoring, reliability, performance, and CLI safety.

• Observability: --verbose flag, test output logging, captured partial credit output, specific exception handlers, integrity checks in awb run
• Scoring: SECURITY_METHODOLOGY capability, signed lint delta, removed hardcoded METRIC_WEIGHTS, timeout calibrator can increase, leaderboard uses per-task aggregate scoring
• Reliability: KeyboardInterrupt handling, load_single None guard, find_incomplete_run scans all _runN dirs, 600s setup timeout, return_exceptions in gather, finally cleanup
• Performance: bare-clone cache (~/.cache/awb/clones/), cached RunEnvironment/adapter, schema cache
• CLI safety: confirmation prompt (--yes), quickstart is env-only check, resolved paths, check_available guard for stubs

1.0.6 (2026-04-03)

• Add trustme to 4 real httpx repo tasks (BF-003, BF-011, BF-013, FA-005)

1.0.5 (2026-04-02)

• Add trio to 16 httpx-based tasks (fixes silent pytest crash on Python 3.13+)

1.0.4 (2026-04-01)

• Fix 4 verification bugs (FA-010, RF-012, CR-007, BF-003)

Older releases

See CHANGELOG.md for the full history (v1.0.0, v0.5.x, v0.4.x, v0.3.x, v0.2.x, v0.1.0).

Links

• Methodology — Fair comparison principles, metric definitions, related work, known limitations
• Architecture — Module graph, data models, pipeline diagrams
• Contributing — Adding tasks, tools, and submitting results
• PyPI — pip install awb

Citing AWB

If you use AWB in research, cite it via Zenodo. The concept DOI 10.5281/zenodo.20361437 always resolves to the latest release; each release also mints a version-specific DOI listed on the Zenodo record. Machine-readable metadata lives in CITATION.cff and codemeta.json; release process is in docs/zenodo-doi.md.

@software{puspus_awb_2026,
  author    = {Puspus, Xavier},
  title     = {{AWB: AI Workflow Benchmark}},
  version   = {1.4.0},
  year      = {2026},
  month     = may,
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20361437},
  url       = {https://doi.org/10.5281/zenodo.20361437}
}

License

MIT