Skip to content

sjarmak/agent-diagnostics

Repository files navigation

Agent Diagnostics

A behavioral taxonomy, annotation framework, and shareable dataset backend for analyzing why coding agents succeed or fail on benchmark tasks.

11,995 trials. 4 models. 61 benchmarks. 40 failure categories across 11 dimensions.

What this does

Coding agents pass benchmarks for the wrong reasons and fail them for the wrong reasons. Pass/fail scores hide reward hacking, flawed tests, and lucky patches. This project extracts structured signals from agent trajectories, classifies failure modes, and provides a queryable dataset backend so you can actually understand what happened.

Install

pip install agent-diagnostics

Quick start: label your own traces

Point the tool at a directory of agent trial outputs. Each trial needs a result.json (and optionally a trajectory.json) in its own directory:

my-runs/
  trial-001/
    result.json          # must have: task_id or task_name, reward/score
    trajectory.json      # optional: list of agent steps with tool calls
  trial-002/
    result.json
    agent/
      trajectory.json    # also checks agent/ subdirectory
  ...

Then run the pipeline:

# Step 1: Extract signals from all trial directories into JSONL
agent-diagnostics ingest --runs-dir my-runs/ --output data/signals.jsonl

# Step 2: Classify failure modes (heuristic rules, instant)
agent-diagnostics annotate --signals data/signals.jsonl --output data/heuristic.json

# Step 3: LLM-assisted classification (reads actual trajectories)
agent-diagnostics llm-annotate --signals data/signals.jsonl --output data/llm.json \
    --sample-size 50 --model haiku --backend batch

# Step 4: Query results with SQL
agent-diagnostics query "SELECT model, count(*) as n,
  round(avg(CASE WHEN passed THEN 1.0 ELSE 0.0 END)*100, 1) as pass_rate
  FROM signals GROUP BY model ORDER BY pass_rate DESC"

What result.json looks like

The tool reads standard benchmark harness output. At minimum:

{
  "task_name": "django__django-16527",
  "reward": 1.0,
  "agent_info": { "name": "claude-code" },
  "started_at": "2026-01-15T10:00:00Z",
  "finished_at": "2026-01-15T10:04:32Z"
}

Works out of the box with SWE-bench, OpenHands, and similar harnesses that write result.json per trial.

What you get back

signals.jsonl — one row per trial with 31 structured fields:

{
  "trial_id": "726c23ceb1ce7cf2...",
  "task_id": "django__django-16527",
  "model": "claude-sonnet-4-6",
  "reward": 1.0,
  "passed": true,
  "total_turns": 57,
  "tool_calls_total": 32,
  "search_tool_calls": 8,
  "edit_tool_calls": 12,
  "unique_files_read": 5,
  "unique_files_edited": 2,
  "duration_seconds": 272.0,
  "exception_crashed": false,
  "tool_call_sequence": ["read_file", "search", "edit_file", "..."],
  "..."
}

Included dataset

The repo ships a Parquet export of 11,995 trials in data/export/ (~1.5 MB):

Model Trials Pass rate
Claude Haiku 4.5 6,443 79.1%
Claude Sonnet 4.6 4,564 73.2%
Claude Opus 4.6 677 84.5%
Claude Opus 4.5 253 71.9%

Query it immediately after cloning:

agent-diagnostics query "SELECT model, count(*) as trials FROM signals GROUP BY model"

# Or load directly
python3 -c "import pandas as pd; print(pd.read_parquet('data/export/signals.parquet').describe())"

Pre-built queries

Five analysis queries are in docs/queries/:

agent-diagnostics query "$(cat docs/queries/per_model_outcomes.sql)"
agent-diagnostics query "$(cat docs/queries/benchmark_model_matrix.sql)"
agent-diagnostics query "$(cat docs/queries/annotation_cooccurrence.sql)"
agent-diagnostics query "$(cat docs/queries/tool_sequence_patterns.sql)"
agent-diagnostics query "$(cat docs/queries/eval_subset_export.sql)"

Export your own Parquet

agent-diagnostics export --format parquet --out data/export/

Produces zstd-compressed Parquet with native list<string> columns, plus MANIFEST.json with schema version, row counts, SHA256 checksums, and source commit.

Schema introspection

agent-diagnostics db schema --format markdown
agent-diagnostics db schema --format json

Taxonomy

40 categories across 11 behavioral dimensions (v3):

Dimension Categories Examples
Retrieval 3 retrieval_failure, query_churn, context_window_overflow
ToolUse 4 wrong_tool_selection, tool_argument_error, tool_misinterpretation
Reasoning 3 decomposition_failure, incorrect_root_cause, overconfident_diagnosis
Execution 5 edit_verify_loop_failure, syntax_error_loop, incomplete_implementation
Environment 4 exception_crash, rate_limited_run, environment_mismatch
Faithfulness 2 task_misunderstanding, scope_drift
Metacognition 5 premature_submission, excessive_exploration, sunk_cost_persistence
Integrity 2 test_file_modification, reward_hacking
Safety 3 data_exfiltration_attempt, sandbox_escape, destructive_operation
Strategy 6 success_via_code_nav, success_via_semantic_search, success_via_decomposition
Observability 3 insufficient_provenance, task_ambiguity, unreproducible_result
from agent_diagnostics import load_taxonomy, valid_category_names

taxonomy = load_taxonomy()
names = valid_category_names()

Annotation pipeline

Heuristic annotation

23 rule-based classifiers that fire on signal patterns (e.g., retrieval_failure when search calls = 0 and files read = 0):

agent-diagnostics annotate --signals data/signals.jsonl --output heuristic.json

LLM annotation

Reads actual trajectories and classifies with Claude. Supports claude-code, api, and batch backends. For anything beyond a handful of trials, prefer --backend batch (Message Batches API): 50% cheaper than the synchronous API and not subject to per-minute rate limits. Use claude-code (the default) only for small interactive samples where you don't want to set ANTHROPIC_API_KEY:

agent-diagnostics llm-annotate --signals data/signals.jsonl --output llm.json \
    --sample-size 50 --model haiku --backend batch

Ensemble (heuristic + classifier)

Two-tier: heuristic rules for structural categories, trained classifier for learned categories. blend merges corpus-scale heuristic labels with the sampled LLM labels into a larger training set (--max-heuristic-samples caps the heuristic-only share; dropped trials are reported, never silent):

agent-diagnostics blend --heuristic heuristic.json --llm llm.json --output blended.json
agent-diagnostics train --labels blended.json --signals data/signals.jsonl --output model.json
agent-diagnostics ensemble --signals data/signals.jsonl --model model.json --output ensemble.json

train cross-validates each category (stratified k-fold) and stores held-out eval_f1 and cv_ece in the model; ensemble only trusts categories whose held-out F1 clears --min-f1 (optionally also gating on calibration with --max-ece).

Annotation store

All annotation writers can route through a shared AnnotationStore that enforces primary key uniqueness, atomic writes, and version consistency:

agent-diagnostics annotate --signals data/signals.jsonl --output heuristic.json \
    --annotations-out data/annotations.jsonl

agent-diagnostics ensemble --signals data/signals.jsonl --model model.json \
    --output ensemble.json --annotations-out data/annotations.jsonl

The store uses PK (trial_id, category_name, annotator_type, annotator_identity, taxonomy_version) so multiple annotators (heuristic, LLM, classifier, ensemble, human) can label the same trial without collision.

Platform note: concurrent-writer safety relies on fcntl.flock, which is Unix-only (Linux/macOS). On Windows the store works for single-process use but does not guard against concurrent writers.

Data formats

signals.jsonl

One JSON object per line. 31 fields per trial including trial_id (stable SHA256-based), model, benchmark, reward, pass/fail, tool call counts/sequences, files read/edited, duration, error counts, and patch size.

annotations.jsonl

Narrow-tall schema — one row per (trial, category, annotator):

Column Description
trial_id SHA256-based stable identifier
category_name Taxonomy category (e.g., retrieval_failure)
confidence 0.0 to 1.0
evidence Free-text explanation
annotator_type heuristic, llm, classifier, ensemble, or human
annotator_identity e.g., heuristic:rule-engine, llm:haiku-4
taxonomy_version e.g., 3.0.0
annotated_at ISO 8601 timestamp

CLI reference

agent-diagnostics ingest           Ingest trial directories into signals.jsonl
agent-diagnostics extract          Extract signals from a single trial directory
agent-diagnostics annotate         Heuristic annotation
agent-diagnostics llm-annotate     LLM-assisted annotation
agent-diagnostics blend            Merge heuristic + LLM labels into a training set
agent-diagnostics train            Train per-category classifiers
agent-diagnostics predict          Predict with trained classifier
agent-diagnostics ensemble         Two-tier ensemble annotation
agent-diagnostics report           Generate Markdown + JSON report
agent-diagnostics calibrate        ECE, Brier, reliability diagrams vs a reference
agent-diagnostics agreement        Pairwise inter-annotator Cohen's kappa per category
agent-diagnostics validate         Validate annotations against schema
agent-diagnostics query            Run SQL against the dataset (DuckDB)
agent-diagnostics export           Export to Parquet with MANIFEST.json
agent-diagnostics manifest refresh Rewrite manifests.jsonl
agent-diagnostics db schema        Inspect table schemas

Calibration metrics

Calibration asks a different question than agreement: not do the annotators pick the same categories? but when the annotator says a category is present with confidence 0.8, is it present 80% of the time? Three proper scoring rules are reported by agent-diagnostics calibrate:

  • Expected Calibration Error (ECE)[0, 1], lower is better. Predictions are bucketed into equal-width bins on [0, 1] (default 10). For each bin, we compute the gap between the mean confidence and the observed accuracy; ECE is the sample-size-weighted average of those gaps. An ECE of 0.0 means the confidences match empirical frequency exactly. An ECE of 0.49 means the model's confidences are off by roughly 49 percentage points on average.
  • Brier score[0, 1], lower is better. Mean squared error between emitted confidence and the binary outcome. Unlike ECE, Brier penalises individual-sample error (not just per-bin averages), so it disambiguates "overconfident on half, underconfident on the other half" from "well calibrated overall".
  • Reliability diagram — per-bin counts, mean confidence, and observed accuracy. Serialised as JSON for downstream plotting. A perfectly calibrated annotator's diagram lies on the y = x diagonal.

The Markdown report also includes a direction arrow (>> overconfident, << underconfident, = well calibrated within 1pp).

How to run

agent-diagnostics calibrate \
  --predictor data/heuristic.json \
  --golden-dir tests/fixtures/golden_corpus/ \
  --output-dir reports/calib/

--predictor is the annotation file whose emitted confidences are being scored. --reference takes a second annotation file as ground truth, or --golden-dir composes the golden corpus's per-trial expected_annotations.json files into the reference. Outputs are calibration.md and calibration.json in --output-dir.

Calibration is only meaningful on predictors that emit per-category confidences. Legacy observatory-annotation-v1 files (produced before the annotator started emitting confidence) default every present category to 1.0 on read, which makes every assigned category look maximally overconfident. Run the pipeline with the current LLM annotator to produce a v2 file before interpreting the ECE / Brier output.

Scope

These metrics describe how well calibrated the annotator's confidences are. They do not fix miscalibration — post-hoc calibration (temperature scaling, Platt, isotonic regression) is a separate concern and is not performed here.

Inter-annotator agreement

Agreement asks the question calibration doesn't: do independent annotators assign the same categories at all? When several annotators (heuristic, LLM, classifier, ensemble) write to the same narrow-tall store, agreement computes pairwise Cohen's kappa per category over the trials both annotators labeled — kappa corrects raw agreement for the agreement expected by chance from each annotator's base rates:

agent-diagnostics agreement \
  --annotations data/annotations.jsonl \
  --output-dir reports/agreement/

Outputs agreement.md and agreement.json with per-pair shared-trial counts and per-category contingency tables. kappa = 1 is perfect agreement, 0 is chance level, negative is systematic disagreement; categories where both annotators are constant are reported as undefined rather than inflated.

Contributing

We welcome contributions of agent trace data, new benchmark integrations, taxonomy refinements, and annotation tooling. If you're building evaluation infrastructure for coding agents, we'd love to talk.

License

Apache-2.0

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors