Overview
Observability in Coevolved is built on structured events emitted during execution.
You use tracing to answer questions like:
- Which steps ran (and in what order)?
- Where did time go?
- What did the LLM return (text vs tool calls)?
- Which tool executed, with what args, and what result?
Tracing configuration
Start by deciding what you want to capture:
- step lifecycle events (start/end/error)
- snapshots (commonly: input snapshot only on errors)
- LLM prompt text (often off by default for privacy and cost)
If you need richer debugging, enable additional snapshots in a controlled environment.
Sinks receive events; formatters turn them into log lines or structured payloads.
Common approaches:
- local dev: console sink with a compact formatter
- staging: JSON lines to a file (JSONL)
- production: forward events into your logging/metrics pipeline
Here’s a simple JSONL setup:
from coevolved.base import (
ConsoleSink,
JSONLSink,
Tracer,
make_verbose_jsonl_serializer,
set_default_tracer,
)
serializer = make_verbose_jsonl_serializer(
include_none=True,
add_timestamp_iso=True,
)
tracer = Tracer(
sinks=[
ConsoleSink(),
JSONLSink("coevolved.trace.jsonl", serializer=serializer),
]
)
set_default_tracer(tracer)
Redaction and privacy
Treat prompts, tool args, and snapshots as potentially sensitive.
Recommended defaults:
- Don’t log full prompt text unless you explicitly need it
- Capture snapshots on errors only
- Implement a redactor for hashing/logging when state can contain secrets
If you handle user data, ensure your logging and retention policies match your security and compliance requirements.
Next steps
