Datasets:
agentlens stringlengths 2.44k 65.2k | agentlens_exported bool 1
class | cost_usd float64 0 0.27 | ended float64 1.79B 1.79B | episode_id stringlengths 44 70 | error stringclasses 3
values | family stringclasses 2
values | label stringclasses 11
values | label_confidence float64 0.4 1 ⌀ | label_evidence stringclasses 746
values | model stringclasses 2
values | outcome stringclasses 4
values | run_kind stringclasses 3
values | runaway_reason stringclasses 1
value | started float64 1.79B 1.79B | task_id stringclasses 95
values | trace_id stringlengths 26 26 | turns stringlengths 2 40.7k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
{"format": "agentlens/v1", "trace_id": "01M2V1HGBP5E2PMR0T7B1A6462", "name": "episode:tools/ops_hard_triage_all_open", "started_at": "2026-09-18T19:57:50.838728Z", "ended_at": "2026-09-18T19:58:44.478340Z", "duration_ms": 53639.612, "status": "ok", "attributes": {"episode_id": "tools-ops_hard_triage_all_open-01M2V1HGBP... | false | 0.1573 | 1,789,761,524.479923 | tools-ops_hard_triage_all_open-01M2V1HGBPRH20S4FPY287T1SD | null | tools | null | null | null | claude-sonnet-5 | pass | nightly | null | 1,789,761,470.838764 | ops_hard_triage_all_open | 01M2V1HGBP5E2PMR0T7B1A6462 | [{"assistant": [{"signature": "EpICCpABCBEYAipAv/szOYMsh4ksXCPVwvCpVy0OSKqJ/JohhZIbLEUYDzuCGlh5QfGIyIVzY77jtDqlih8kEkZW9UZ3v/ENVqgu5jIPY2xhdWRlLXNvbm5ldC01OABCCHRoaW5raW5nWiRjNjcxNGRlMy1lYjJkLTQwZWItYWYxNi0zMDI3YjliMmRjOGKoAcCvttUGEgz8o/D8SvSNbhmAx5waDOotRlMs308PkiyPyyIwAijQxoMvM2t1VxL1LWCdvJ4bp+PJxYl2v0kqtZ9HrJdgkFVCu... |
{"format": "agentlens/v1", "trace_id": "01M33VEYPZGP1YQWT7YHYX8ZR1", "name": "episode:tools/cal_spill_to_next_day", "started_at": "2026-09-22T06:04:45.663512Z", "ended_at": "2026-09-22T06:04:56.666540Z", "duration_ms": 11003.028, "status": "ok", "attributes": {"episode_id": "tools-cal_spill_to_next_day-01M33VEYPZVWW4SZ... | false | 0.0273 | 1,790,057,096.667403 | tools-cal_spill_to_next_day-01M33VEYPZVWW4SZMK7APAF7KS | null | tools | null | null | null | claude-sonnet-5 | pass | nightly | null | 1,790,057,085.663557 | cal_spill_to_next_day | 01M33VEYPZGP1YQWT7YHYX8ZR1 | [{"assistant": [{"signature": "EpsCCpABCBIYAipAcAVbJfiYjr/RsOjOgMTkS1OuZBfBG3foD7c9vgUM4IyVm4HgcmbPiZc9C+NFGqyOJT21fMeD3/Th7hNRAA0XrTIPY2xhdWRlLXNvbm5ldC01OABCCHRoaW5raW5nWiRjNjcxNGRlMy1lYjJkLTQwZWItYWYxNi0zMDI3YjliMmRjOGKoAf+0yNUGEgy4MUcSTiYQ4RhBmRcaDPhPOkcbnyPbH+qzLyIw/HU9InqauAzUSY0N0Jkus48RHVf2+Uid20A4IOSgJSiM1TgyC... |
{"format": "agentlens/v1", "trace_id": "01M33NVDMFW0CZY5E6WGC7TWK3", "name": "episode:tools/cal_hard_timezone_spill", "started_at": "2026-09-22T04:26:42.703718Z", "ended_at": "2026-09-22T04:26:50.514911Z", "duration_ms": 7811.193, "status": "ok", "attributes": {"episode_id": "tools-cal_hard_timezone_spill-01M33NVDMFPYH... | false | 0.016 | 1,790,051,210.516419 | tools-cal_hard_timezone_spill-01M33NVDMFPYHWZ4HEDDVT3K4B | null | tools | null | null | null | claude-sonnet-5 | pass | nightly | null | 1,790,051,202.70378 | cal_hard_timezone_spill | 01M33NVDMFW0CZY5E6WGC7TWK3 | [{"assistant": [{"signature": "ErsCCpABCBIYAipAKtUQU2yu2kIdZiClRJB4l0ExsDcXcY+6NDgc353upnmNKKUUx2WUyBNrjAKtBg2b9qMIioJjXbTYLlc0hCcgxjIPY2xhdWRlLXNvbm5ldC01OABCCHRoaW5raW5nWiRjNjcxNGRlMy1lYjJkLTQwZWItYWYxNi0zMDI3YjliMmRjOGKoAYSHyNUGEgza1UMkAYi0dyQA9KsaDIxoZNuJn+lpHS1MaSIw5JylmQs6mxItGk53dKu2836faT1XKHLHG1F45kLvWpgNjfYQY... |
{"format": "agentlens/v1", "trace_id": "01M33YJ7AY94MRHY8525Y8J5TG", "name": "episode:tools/air_stack_disruption", "started_at": "2026-09-22T06:58:58.526412Z", "ended_at": "2026-09-22T06:59:25.759448Z", "duration_ms": 27233.036, "status": "ok", "attributes": {"episode_id": "tools-air_stack_disruption-01M33YJ7AYM69KZSZT... | false | 0.0903 | 1,790,060,365.760856 | tools-air_stack_disruption-01M33YJ7AYM69KZSZTHEWJRYY1 | null | tools | null | null | null | claude-sonnet-5 | pass | nightly | null | 1,790,060,338.526435 | air_stack_disruption | 01M33YJ7AY94MRHY8525Y8J5TG | [{"assistant": [{"signature": "EpwCCpABCBIYAipALmZS3D2AQ/ygF4WxRLHnP17xkBYn9SfajDlTeGXC4tl3FjP6Fgd4+5hKjSalTVaqeDojmmh243bFEzx0AFroOjIPY2xhdWRlLXNvbm5ldC01OABCCHRoaW5raW5nWiRjNjcxNGRlMy1lYjJkLTQwZWItYWYxNi0zMDI3YjliMmRjOGKoAbPOyNUGEgwVafUQSSK6wO+JqoMaDD6AZkuCCEmPIMMmUCIwOslLbzg/tyMjhfaXMPm5RPeOy02/hJlejlWywglTXqkAsAPGA... |
"{\"format\": \"agentlens/v1\", \"trace_id\": \"01M388NY6Z2S547MVM9KHJB2WJ\", \"name\": \"episode:to(...TRUNCATED) | false | 0.1103 | 1,790,205,202.310813 | tools-retail_stack_ticket_queue-01M388NY6ZPZR1WFKPZY9BGB1M | null | tools | null | null | null | claude-haiku-4-5-20251001 | pass | nightly | null | 1,790,205,163.743926 | retail_stack_ticket_queue | 01M388NY6Z2S547MVM9KHJB2WJ | "[{\"assistant\": [{\"citations\": null, \"text\": \"I'll work through the entire support ticket que(...TRUNCATED) |
"{\"format\": \"agentlens/v1\", \"trace_id\": \"01M341JTQVJZ2GSBHZTXAT27F3\", \"name\": \"episode:to(...TRUNCATED) | false | 0.2351 | 1,790,063,567.983746 | tools-retail_stack_ticket_queue-01M341JTQVB3J88C5VSYSFE70Q | null | tools | null | null | null | claude-sonnet-5 | pass | nightly | null | 1,790,063,504.123161 | retail_stack_ticket_queue | 01M341JTQVJZ2GSBHZTXAT27F3 | "[{\"assistant\": [{\"signature\": \"ErACCpABCBIYAipA6xAVaLq67ywjJGGKQxnpAHgxZX0mLrOiZiN0F4NdP8RcOt7(...TRUNCATED) |
"{\"format\": \"agentlens/v1\", \"trace_id\": \"01M2EX2FPV229WH1VWYHZ46VFZ\", \"name\": \"episode:to(...TRUNCATED) | false | 0.0496 | 1,789,354,148.323521 | tools-cal_hard_three_way_two_hours-01M2EX2FPVBQDXB1HXK7TADN8H | null | tools | null | null | null | claude-sonnet-5 | pass | calibration | null | 1,789,354,131.163227 | cal_hard_three_way_two_hours | 01M2EX2FPV229WH1VWYHZ46VFZ | "[{\"assistant\": [{\"signature\": \"EpMDCpABCBEYAipAoGujjPSBeLjFzdRt8SGAoEaOjHnujsgmsNvTc1e27eQ50wT(...TRUNCATED) |
"{\"format\": \"agentlens/v1\", \"trace_id\": \"01M33PKBKHRD954PYNN5YR8XY4\", \"name\": \"episode:to(...TRUNCATED) | false | 0.1114 | 1,790,052,038.605076 | tools-retail_stack_ticket_queue-01M33PKBKH9Y3EKW7SNB1H7YZJ | null | tools | misread_spec | 0.75 | "Turn 4: agent refused TK-1 ceramic refunds, asking Maya to name items, though ticket named 'every c(...TRUNCATED) | claude-haiku-4-5-20251001 | fail | nightly | null | 1,790,051,987.057986 | retail_stack_ticket_queue | 01M33PKBKHRD954PYNN5YR8XY4 | "[{\"assistant\": [{\"citations\": null, \"text\": \"I'll start by fetching the support ticket queue(...TRUNCATED) |
"{\"format\": \"agentlens/v1\", \"trace_id\": \"01M2F0S3DPD1BP76T94Q10W02H\", \"name\": \"episode:to(...TRUNCATED) | false | 0.0069 | 1,789,358,023.21837 | tools-retail_refund_pending_is_cancel-01M2F0S3DM7Q7W48EJCRMJ339T | null | tools | null | null | null | claude-haiku-4-5-20251001 | pass | calibration | null | 1,789,358,017.974986 | retail_refund_pending_is_cancel | 01M2F0S3DPD1BP76T94Q10W02H | "[{\"assistant\": [{\"citations\": null, \"text\": \"I'll help you with this. Let me first fetch the(...TRUNCATED) |
"{\"format\": \"agentlens/v1\", \"trace_id\": \"01M340GJ470W6KEEGWW31C7N7A\", \"name\": \"episode:to(...TRUNCATED) | false | 0.0104 | 1,790,062,390.040405 | tools-cal_hard_three_way_two_hours-01M340GJ47YVH20KNWJ64H9P27 | null | tools | misread_spec | 0.72 | "Turn 1: booked mon 11 for 2h (hours 11-12) though Pat is busy hour 12-13; agent mis-computed overla(...TRUNCATED) | claude-haiku-4-5-20251001 | fail | nightly | null | 1,790,062,381.191051 | cal_hard_three_way_two_hours | 01M340GJ470W6KEEGWW31C7N7A | "[{\"assistant\": [{\"citations\": null, \"text\": \"I need to check the calendars for you, Pat, and(...TRUNCATED) |
End of preview. Expand in Data Studio
AgentLens Failure Corpus
Nightly, unattended runs of tool-using agents against fixed tasks. Every episode is a full trace, and every failure is labeled with a fixed taxonomy. Tasks are scored by deterministic state or equality checkers, with no LLM judges, so an outcome can change only because of the model or its serving, never the benchmark.
- Code, dashboard and paper draft: https://github.com/agentlens-sdk/agentlens-failure-corpus
- Live dashboard: https://agentlens-sdk.github.io/agentlens-failure-corpus/
- Updated: weekly snapshot of a corpus that grows every night.
Design
- Paired: every queued task runs once on each agent model (
claude-sonnet-5,claude-haiku-4-5-20251001) in the same run. - Repeated: coding tasks run several times per night, and tool tasks are sampled with replacement, so the same prompt appears many times per model per night.
- Families:
tools(policy-following over mock retail, airline, incident-ops and calendar APIs; checked against the backend's final state) andflakiness(short coding, bug-fix, refactor and simulation prompts; checked against fixed cases). - Hard by construction: tasks that no listed model fails are retired from the nightly queue. Pass rates on the active set therefore describe behaviour at the edge of these models' ability, not general capability.
Fields
One row per episode.
| field | type | meaning |
|---|---|---|
episode_id |
string | <family>-<task_id>-<ULID> |
trace_id |
string | ULID of the root span |
family, task_id |
string | task identity |
model |
string | requested agent model |
outcome |
string | pass, fail, runaway (hit a turn, token or time cap), error (the episode itself errored) |
run_kind |
string | nightly, calibration or aborted. Use nightly for pass rates. |
label |
string | failure label, null for passes |
label_confidence, label_evidence |
double, string | labeler's confidence and a short cited reason |
turns |
JSON string | per turn: assistant content, usage, stop_reason, served_model, request_id, tool results |
agentlens |
JSON string | agentlens/v1 span tree: agent root, one llm span per turn, one tool span per call |
started, ended |
double | Unix seconds |
cost_usd |
double | agent-side API cost of the episode |
error, runaway_reason |
string | present when relevant |
served_model and request_id are recorded from 2026-09-15 on. Episodes with zero turns never reached the
model (a network failure) and should be excluded from behavioural analysis.
Labels
Coding failures are labeled from the checker by rerunning the submission: wrong_output, crashed,
timed_out. Tool failures are labeled by Claude Opus 5 from the trace plus a replay diff against the
reference solution: misread_spec, premature_stop, tool_misuse, wrong_target, hallucinated_tool,
loop, gave_up, environment_error, other. A reply without a valid label is recorded as
unclassified. In a 2026-09-14 audit, Haiku 4.5 agreed with Opus 5 on 45 of 70 tool failures (64%), and Opus
was right on every disagreement reviewed.
Limitations
The tasks were selected adversarially against these two models, both agents are Claude models at API-default sampling, the environments are small mock APIs, and tool labels come from a single model labeler. The time span is short, so night-to-night drift is not yet measurable at useful power.
License
Data: CC-BY-4.0. Code: MIT.
- Downloads last month
- 76