Execution paths
See the exact order of steps, branches, retries, tools, and handoffs that produced an answer.
See exactly how your AI behaves, at every step
Trace prompts, workflows, agents, tools, and sub-agents in one place. Follow every execution path, inspect latency and usage, and understand what happened before your users report it.
TRACE
ANALYZE
IMPROVE
agent.public_request
req_01JX9F7M4K2
COMPLETED
1.82 s
Latency
3,842
Tokens
$0.014
Cost
EXECUTION PATH
agent.invoke
1.82 s
tool.search
428 ms
provider.llm
1.21 s
Every span stays connected to the request, model, workspace, and final cost.
Trace every AI interaction from request to result
Open one run and follow the complete execution path. Inspect model calls, workflow steps, tool actions, sub-agent handoffs, outputs, errors, duration, usage, and cost without stitching evidence together by hand.
Prompt, workflow, and agent runs in one view
Use the same trace language across every way your team ships AI. Compare behavior across models and providers without changing how you investigate a run.
Prompts
Inspect model choice, input and output, duration, tokens, cost, and the execution metadata behind each completion.
provider.llm → completion
Workflows
See branch decisions, step order, parallel work, retries, failures, and cost all the way down to an individual task.
build → enqueue → execute
Agents
Follow the parent run through tool use, sub-agent calls, model work, and the final response with one connected context.
agent → tool → sub-agent
Across the models you already use
Connect performance, reliability, and cost
Move from one trace to the bigger pattern. Monitor latency, errors, token usage, model behavior, and spend, then return to the exact run behind an outlier.
Find slow providers, models, and execution paths
Spot errors and investigate the run behind the trend
Attribute usage and cost to the work that created it
Production health
Last 24 hours
LIVE
1.82 s
Median latency
18.4K
Tokens used
$42.81
Total cost
2.1%
Failed runs
LATENCY BY RUN
p95 3.46 s
One request. Two observability layers. Full context.
Fetch Hive keeps customer-facing product traces durable while OpenTelemetry provides selective cross-service correlation for engineering. Shared identifiers connect the two without pretending sampled telemetry is your audit record.
Durable product traces
The customer-facing waterfall, completion detail, duration, usage, and cost views are backed by durable product trace rows.
trace_id trc_01JX9F7M
run_source fetch_chat
target_kind saved_agent
OpenTelemetry correlation
Selective spans connect Rails, Rust, provider round trips, workflow launches, tool calls, and completion handoffs across service boundaries.
otel_trace_id 6f3b…a912
fetchhive.request_id req_01JX
service.name rust-api
SHARED CONTEXT
Keep the evidence. Leave sensitive material out.
Observability should make production AI easier to trust, not create a second data problem. Fetch Hive separates durable product evidence from selective telemetry and bounds the payloads it serves.
Debug AI with the evidence in front of you
Build, ship, and improve prompts, workflows, and agents with tracing that stays connected to real production outcomes.