Agent Reliability

Your monitoring has a blind spot

Traditional observability records infrastructure signals: latency, error rates, and token counts. It cannot see an agent's output get worse, and the eval scores layered on top stop at telling you a number moved. When your agent degrades in production, Tessary tells you which change caused it, in time to fix it before it does damage.

APM (Datadog, Grafana)
Latency, error rates, token counts, HTTP status codes. Infrastructure health, no view of output quality.
LLM observability (Langfuse, Braintrust)
Traces, prompts, and eval scores you assemble yourself. They tell you a score moved, not which change moved it.
Agent reliability (Tessary)
Reads the trend over graders, then traces a degradation to the causing change. Pulls from your existing Langfuse or Braintrust traces, no re-instrumentation.
The evidence

What the published data shows

A NeurIPS benchmark study and Datadog's telemetry point at the same gap: the failures that matter most don't appear on an infrastructure dashboard.

41–86.7%

Task failure rates on benchmark tasks across seven popular multi-agent frameworks

MAST, NeurIPS 2025

60%

Of LLM call errors in February 2026 were caused by exceeded rate limits, the failure class most likely to be masked by retries

Datadog, State of AI Engineering 2026

The gap

Full observability. Invisible failures.

In April 2026, engineers running Claude Code with full observability tooling had no automated signal that behavior had degraded. Anthropic's postmortem records that infrastructure modifications altered agent behavior for weeks before user complaints surfaced the problem. Latency and error rates showed nothing unusual throughout.

Anthropic postmortem, Apr 2026

Weeks

Agent behavior degraded before any monitoring alert fired

Zero

Infrastructure alerts during the full degradation period

Three tiers

What each layer answers

Datadog and Grafana sit a tier below both columns: infrastructure health, with no view of output at all. LLM observability holds the traces and tracks the scores. Agent reliability starts when a score moves and ends at the change that moved it. The layers are complementary, not competing.

LLM observability

Langfuse, Braintrust, and eval platforms

  • Traces, prompts, and outputs stored and searchable
  • Eval scores tracked over time
  • Latency, token counts, and cost per call
  • Degradation read as a trend across graders
  • A score drop traced to the change that caused it
  • Causes with no diff: provider updates, upstream agent shifts
Tessary

Agent reliability

Tessary, on top of your existing traces

  • Traces, prompts, and outputs stored and searchable
  • Eval scores tracked over time
  • Latency, token counts, and cost per call
  • Degradation read as a trend across graders
  • A score drop traced to the change that caused it
  • Causes with no diff: provider updates, upstream agent shifts
Blind spots

Failures your stack won't flag

These failure classes produce no HTTP error, no exception trace, and no latency spike. They are invisible to APM tools by design.

Hallucinated tool calls

The model invokes a nonexistent or incorrect function. The request returns HTTP 200 with no error trace. Infrastructure metrics look healthy while the agent output is wrong.

Output quality decay

After a model version update, accuracy drops gradually. No threshold alert fires because no infrastructure metric changed. The degradation is invisible until users report it.

Rate-limit retry masking

A retry succeeds with a fallback model but delivers a degraded response. The span records a successful retry, not the quality loss. No metric captures the difference.

Multi-agent coordination gaps

A sub-agent silently skips a step. The orchestrator proceeds with incomplete context. No exception fires. The failure is visible only in the final output.

How it works

Four stages of reliability

Tessary layers on top of the traces you already have. The flow ends at the cause, not at an alert that a number moved.

  1. Connect your trace source

    Link Tessary to Langfuse or Braintrust as an upstream source, or send traces through the OTLP receiver. Tessary pulls a chosen slice of your already-instrumented traces on demand. No instrumentation changes required to start.

  2. Detect degradation as a trend, not a threshold

    Tessary scores traces against graders and reads the trend across verdicts over time, not any single score. An imperfect grader misfires at a stable rate, so a failure rate that jumps from 3% to 11% signals a real change in the agent. Detection holds up before your graders are perfect.

  3. Trace the drop to the change that caused it

    A failing grader names the quality dimension that broke. Tessary walks that dimension back through the agent's chain to the change responsible: a prompt edit, a tool update, a dependency bump, an upstream agent's output shift, or a model provider update that never touched your repo. When the trail hits a gap in the traces, that gap becomes the next node to instrument, and the next hunt is faster.

  4. Route the finding, cause attached

    The alert lands in Slack, Sentry, PagerDuty, Linear, or a webhook, and it names the cause rather than just the score that moved. The same attribution engine can also run before a merge and post a risk report on the PR, for the causes that do ship through one.

Questions

Answered, plainly.

Three tiers do different jobs. APM tools like Datadog track infrastructure: HTTP status codes, latency, token counts. LLM observability platforms like Langfuse and Braintrust capture prompts, outputs, and traces, and track eval scores you assemble yourself; they tell you when a score moved. Agent reliability is the layer above that: Tessary reads the trend over graders to confirm a real degradation, then traces it back to the change that caused it, including changes that never touched your repo.
APM tools like Datadog and Grafana surface rate-limit errors, timeouts, and exception traces. They cannot evaluate whether a tool call returned a semantically correct result, whether output quality has decayed compared to a prior baseline, or whether a rate-limit retry delivered a worse response than the original request. Those failures require a grading layer over the agent's actual behavior.
Semantic agent failures are production errors that don't produce an HTTP error, exception, or monitoring alert. Infrastructure metrics stay healthy, but the agent's output is wrong or degraded. Common examples: hallucinated tool calls that return HTTP 200 with incorrect data, output quality decay after a model version update, and multi-agent coordination failures where sub-agents silently skip steps.
Keep it. Tessary connects to Langfuse or Braintrust as upstream sources and pulls your already-instrumented traces on demand, with no re-instrumentation. Those platforms stay the system of record for traces and scores; Tessary adds the layer that says which change moved a score. Findings route to Slack, Sentry, Linear, or PagerDuty with the cause attached.
No. Tessary drafts graders from the agent's observed behavior, and you correct them rather than write them from scratch. The intent behind the graders lives outside the agent's code and sharpens over time as bugs are fixed, edge cases are shared, and existing golden datasets are folded in. Some review from the agent owner keeps it honest, far less work than authoring and maintaining evals by hand.
It is the same attribution engine, run before the merge. Tessary reads the PR diff, maps the touched surface to failure patterns it has observed on real traffic, and posts a risk report on the PR; a GitHub Action can gate the merge. One boundary to know: a gate like this only fires on changes that ship through a PR. A model provider update, an upstream agent's shift, or a tool returning different data reintroduces a failure with no deploy at all, which is why attribution runs in production first.
Tessary ingests agent traces via an OTLP receiver in gen_ai OpenTelemetry conventions, or via upstream pull connectors for Langfuse and Braintrust. Production SDKs for TypeScript and Python are available. If you are already instrumented on Langfuse or Braintrust, connecting takes a single configuration step. If you are starting from scratch, the TypeScript or Python SDK adds tracing to your agent in under an hour.
Get started

Find what caused it before users report it

Connect your Langfuse or Braintrust traces in one configuration step, with no instrumentation changes and no credit card, and get from a score drop to the change behind it.