Agent Reliability

Replay every agent failure

Walk a failed production session turn by turn. Find where the decision chain went wrong and connect it to the change that caused it: a code deploy, a prompt edit, a tool that started returning different data, or a model update that never touched your diff.

The shift

Agents fail silently in production

Standard monitoring reports normal latency, normal error rates, and normal throughput while an agent works through a session incorrectly. A tool call at turn 3 feeds context that shapes the decision at turn 7, which reaches turn 9 as a confident wrong answer. No exception fires. No alert triggers.

The cost

Errors are not the failure signal

Datadog measured errors across LLM call spans in February 2026. The errors that do fire are mostly infrastructure noise. The failures that cost you, wrong decisions carried confidently to the final turn, produce no error at all.

5%

of all LLM call spans reported an error in February 2026

Datadog, State of AI Engineering 2026

60%

of those errors were exceeded rate limits, not reasoning or decision failures

Datadog, State of AI Engineering 2026

Two approaches

Dashboards tell you a score moved. Replay shows why.

Trace viewers and score dashboards

What you have today

  • Show every span and every score, once you know where to look
  • Tell you a score moved, not which change moved it
  • The walk from moved score to causing change is manual
  • No lineage connecting a session to the change that was live when it ran
Tessary

Tessary session replay

Navigate session → turn → trace

  • Traces the failing quality dimension back through the session to the suspect turn
  • Each turn shows the full context, tool calls, and returned values the agent had
  • Ties the failure to the change behind it, whether or not it shipped through your repo
  • Works with existing Langfuse or Braintrust instrumentation
How it works

The Session-to-Turn-to-Trace model

Graders score every production session and watch the trend, so a degrading pattern surfaces the sessions worth replaying before a customer reports one. From there, Tessary makes the first pass and you verify.

  1. Open the session timeline

    The session timeline shows all turns in order. Tessary gives every production run a navigable structure: a session holds the full sequence of turns, each with its typed observations and tool calls.

  2. Land on the suspect turn

    Tessary traces the failing quality dimension back through the session and opens the turn where it first broke. You review the context the agent had at that point, retrieved documents, prior tool outputs, and accumulated state, rather than hunting for it turn by turn.

  3. Verify the deviation

    Replay is the verification pass. Confirm whether retrieved context went stale, a tool returned an unexpected schema, or accumulated state led to a wrong decision, with every tool call and the value it returned in front of you.

  4. Trace to the change

    Commit-SHA lineage tags every session to the code version running when it occurred, which covers the code-change case with the diff already attached. When no deploy lines up, Tessary correlates the causes that never touch your repo: a tool that started returning different data, an upstream agent's shift, or a model provider update.

What replay shows

The inputs to each decision

Standard error logs record whether each step completed. Replay records why each decision was made.

Full context at each turn

The retrieved documents, prior tool outputs, and accumulated state the agent had at that turn: the exact inputs to the decision, not just the output.

Tool calls and returned values

Every tool call made at each turn, with the parameters it received and the output that propagated forward to the next decision.

The deviation point

The specific turn where behavior first diverged from expected, visible as a gap between the context available and the output produced.

The change that was live

The commit SHA running when the session occurred covers the deploy case. Tool output shifts and model provider updates are correlated alongside it, so the trail does not end at your repo.

Field observation
If you are running agents in production without span-level tracing, you are operating a black box. You can detect that something failed. You cannot reliably explain why.
MLflow Agent Observability Guide2026 Developer GuideRead the guide
Questions

Answered, plainly.

Agent failure replay is the process of walking through a production agent session turn by turn to find where behavior deviated from what was expected. Each turn shows the agent's full context, the tool calls it made, and the outputs it received, so you can trace how a wrong decision at one step propagated forward to a wrong result.
Standard logs capture whether each step completed. They do not record the context and intermediate state the agent had at each decision point, which is the information needed to understand why the output was wrong. An agent failure often involves no error code at all: the session completes with 200 OK, but the agent made a wrong decision based on stale context or a silently wrong tool output.
Commit-SHA lineage means every agent session in Tessary is tagged with the code version that was running when it occurred. When an incident fires, you can identify the deploy that preceded the failure pattern and examine the diff that introduced the change, without guessing which release was in production at the time. When the pattern lines up with no deploy at all, the cause usually shipped outside your repo, a tool returning different data or a provider model update, and Tessary correlates those changes too.
Yes. Tessary can pull a chosen slice from Langfuse or Braintrust as an upstream source without requiring new instrumentation in your agent. Once the data is in Tessary, it is first-class for session navigation, grader scoring, and commit-SHA lineage. The difference is what happens after a score moves: Langfuse and Braintrust tell you a metric dropped, and Tessary walks the drop back through the session chain to what changed.
Traditional debugging attempts to reproduce the failure locally or in staging. Agent failures are often non-reproducible: non-zero temperature means the same input produces a different output. Replay reconstructs the actual production session that failed, including the exact context and tool call sequence at every turn, rather than attempting a new reproduction run.
After the cause is verified through replay, you can promote the session trace to a grader in Tessary. That grader keeps scoring production sessions, so a recurrence shows up as a trend break within hours, whichever kind of change reintroduced it. The same grader can also run in CI against future deploys, though a CI check only sees changes that ship through a PR; a provider or tool-side shift arrives with no deploy at all.
Get started

Replay your first incident today

Connect Langfuse, Braintrust, or OTLP and read the session that failed before the next one fires. The trail ends at what changed, not at a moved score, and no credit card is needed to start.

Find what broke your agentNo re-instrumentation needed on Langfuse or Braintrust. When a replay trail goes dark, that gap is the next node to instrument, and the next failure is easier to trace.