Strands Agents¶
You need a Strands Agents Agent, Graph or Swarm. Ingesting
an export needs no extra; driving the agent needs the strands extra.
There are two ways in.
Ingest a Strands OTel export¶
Strands emits OpenTelemetry spans through the global tracer provider. Export them as OTLP-JSON
(for example with SpanCapture, as in the GenAI recipe) and
ingest the files. Graph, Swarm and agents-as-tools topologies share one span grammar:
$ calibrated-agent-evaluation ingest fixtures/raw/strands
calibrated-agent-evaluation ingest: 6 ingested, 0 refused, 0 skipped
[ok] fixtures/raw/strands/CLM-4471.otlp.json → CLM-4471 (strands, 5 span(s), 4 tool call(s), owned)
[ok] fixtures/raw/strands/CLM-4473.otlp.json → CLM-4473 (strands, 5 span(s), 4 tool call(s), owned)
[ok] fixtures/raw/strands/CLM-4475.otlp.json → CLM-4475 (strands, 5 span(s), 4 tool call(s), owned)
[ok] fixtures/raw/strands/CLM-4477.otlp.json → CLM-4477 (strands, 5 span(s), 4 tool call(s), owned)
[ok] fixtures/raw/strands/CLM-7720-swarm.otlp.json → CLM-7720-swarm (strands, 2 span(s), 3 tool call(s), owned)
[ok] fixtures/raw/strands/ORD-3310-as-tool.otlp.json → ORD-3310-as-tool (strands, 2 span(s), 2 tool call(s), owned)
The four CLM-* traces are a claims crew built as a Graph, CLM-7720-swarm is a Swarm, and
ORD-3310-as-tool is an agent that calls another agent as a tool.
From Python, normalize(StrandsSession(path=...)) reads one export; pass mode="legacy" for an
export written with Strands' older span attributes.
Let Calibrated Agent Evaluation drive it¶
Give StrandsExecutor a function that builds your topology from a model, and it runs the agent,
records the model calls to a cassette and returns traces. Replayed from the repository's cassette:
trace CLM-4471: ownership=owned, 5 spans
taxonomy cell: {'feature': 'claim_adjudication', 'scenario': 'S2_vacancy_exclusion', 'persona': 'P2_homeowner_absent'}
supervisor parent=root no tools
fraud-scorer parent=supervisor score_fraud_risk
intake parent=supervisor extract_claim_fields
payout parent=supervisor no tools
policy-checker parent=supervisor lookup_policy_exclusions, lookup_policy_exclusions
The code, and how to record your own cassette, is on Route your agent through Calibrated Agent Evaluation.
What you get¶
An owned trace per run: a supervisor span with one child per specialist, each with its tool calls, arguments and results. The taxonomy cell you pass to the task is copied onto the trace.
Limits¶
- Specialists that run in parallel finish in a different order on each run, so compare traces by agent, not by step order.
- A live run calls Bedrock and needs AWS credentials. Replay needs none: the Bedrock model is only built when a cassette miss needs a live call.
- The executor sends
temperature=0.0unless you passtemperature=None; the newest Anthropic models reject an explicit temperature.