Tensortable - Continual Learning Layer for AI Agents

01 agentic.tracing

Understand what your agents are doing

OpenTelemetry-native tracing of agent runs with no vendor lock-in. Capture LLM calls, tool calls, sub-agent behavior, costs, and token metrics with single-line instrumentation.

demo_session800ms$0.0021[]

PreviewspCollapse repeats

208 observations·95 rows·32 in DOM·indent cap: 23·11 runs collapsed·113 rows folded

demo_session16m 48s$0.7104

research_pipeline8m 52s$0.4218

ClaudeAgent.query8m 52s$0.4218

anthropic.messages.create4.30s$0.01421240 → 380helpfulness: 0.92

I'll research OpenTelemetry's AI observability features using a multi-step approach, starting with the GenAI semantic conventions.

Skill35ms

anthropic.messages.create3.70s$0.0098980 → 210

Workflow71ms

local_workflow6m 40s

anthropic.messages.create6m 43s$0.128412400 → 3100hallucination: 0.12relevance: 0.88

Based on the gathered facts, OpenTelemetry's GenAI semantic conventions define spans for LLM calls, embeddings, and tool invocations. The gen_ai.* attribute namespace covers request model, response model, token usage and finish reason.

structured_output_retry2.40s$0.0061820 → 140x4WARNING

{
  "error": "response did not match schema: expected confidence to be number"
}

anthropic.messages.create8.40s$0.02312100 → 640

Agent45ms

analyst1m 24s$0.1861

anthropic.messages.create14.00s$0.03413200 → 890

anthropic.messages.create2.00s$0.0051410 → 120

Bash910msERROR

exit status 1: ModuleNotFoundError: No module named 'scipy'

anthropic.messages.create16.60s$0.04124100 → 1020

anthropic.messages.create6.30s$0.01631500 → 430

Bash106ms

anthropic.messages.create1m 1s$0.02982900 → 760accuracy: 0.95

Agent38ms

writer58s$0.0912

apply_patch400ms

vector_search1.80s

embed_query420ms$0.0003

safety_check210ms

checkpoint2ms

OutputContext windowAsk/Annotate

Context window for

anthropic.messages.create

obs-0000

Session turn

123456789101112

Window by source224 / 8,000 tokens (3%)

System prompt148User turn76

Diff vs turn —

First turn — nothing to compare.

Segments (2)

0System promptsystemYou are a research assistant. Cite every factual claim with t…148

1User turnuserTurn 1: follow-up question from the user.76

Content is stored content-addressed by content_hash. Across these 12 turns that is 91 segment references resolving to 52 unique bodies — a 1.8x reduction versus copying each message into every span.

02 anomaly.detection

Automatically detect hidden failures

Agent-driven investigation to automatically analyze your telemetry in a unified harness, revealing hidden failure modes that you would've missed otherwise.

Anomalies

New detector

Active (4)

Structured-output retry cascadetracking

The agent retried the same structured-output call three or more times because the response failed schema validation.

412

Findings

6

Clusters

25%

Sampled

v3 · system modellast 2026-08-21 09:14

Ungrounded claimdiscovery

The final output states a specific figure, name or date that appears in no tool result in the trace.

138

Findings

4

Clusters

50%

Sampled

v2 · system modellast 2026-08-21 09:14

Dropped constraintdiscovery

A constraint the user stated earlier in the session is absent from the model's context window on a later turn.

74

Findings

3

Clusters

100%

Sampled

v1 · system modellast 2026-08-21 09:14

Silent tool failuretracking

A tool returned an error or empty payload and the agent continued without acknowledging or recovering from it.

221

Findings

5

Clusters

25%

Sampled

v1 · system modellast 2026-08-21 09:14

Disabled (1)

Task left incompletediscovery

The agent's final response does not address every part of the user's request.

0

Findings

0

Clusters

10%

Sampled

v1 · system modellast 2026-08-21 09:14

03 continual.improvement

Continually improve your application

See failures, automatically rootcause and fix them — as one continual self-improving loop for your application.

Harness Learning

← All detectorsStructured-output retry cascadev3

Clusters · 6

confidence emitted as string, not numberopen

231 runs where the model returns confidence as a quoted string. The schema requires number, validation fails, and the call is retried on the same prompt.

231 findings · first seen 2026-08-04

enum casing rejected by the validatorauto-resolved

88 runs returning SCREAMING_CASE for an enum the schema declares lowercase. Prompt v14 restated the casing and the cluster stopped producing findings.

88 findings · first seen 2026-07-29 · auto-resolved 2026-08-14

Findings (verdict = Yes)Runs (all sampled)

VerdictConf.OriginDef. vTraceWhenEvidence

Yes92%livev3tr_9f4c212026-08-21 09:14confidence returned as "0.82"; schema expects number

Yes88%livev3tr_7b1e082026-08-21 08:47third attempt returns the same quoted confidence

Yes84%backfillv2tr_4d90aa2026-08-20 22:03four attempts, identical validation error each time

Yes91%livev3tr_2c55f12026-08-20 17:31confidence emitted as string after a tool result was appended

Yes79%backfillv2tr_e10b732026-08-20 11:56retry budget exhausted; run completed without the field

Yes86%livev3tr_bb30922026-08-20 09:02validator rejected the same key on all four attempts