Most agent observability tools were built for chats that finish in seconds. Yours don’t.

GlassFlow is built for the workflows chat tools weren’t: agents that run for hours and chain hundreds of steps. Crash at step 190? You’ll know exactly where and why.

Get full visibility across all your agent workflows.

Simple short running flows
length 2m
312/mfrontend156/mcheckout
Services 2Latency 2mCost $0.01
Long running and multi framework flows
length 6h
1.3K/mfrontend55/mplanner120/mretriever20/mtools35/mvalidate383/mmemory78/mwrite
Services 7Latency 6hCost $2.14
Always on flows
live
running · 6h 02mModel Claude

Functionalities built for your agent type

Long-horizon visualisation

The waterfall view every tool uses breaks down around 100 events. A 20-minute run with 500 tool calls produces an unreadable waterfall. GlassFlow uses a timeline model with anomaly detection, so you can scan hundreds of steps at a glance and jump straight to the one that matters.

Agent workflow maplength 38m
24mfetch-config41mload-context28menrich-data38mplan-exec8mcodex6mwrite-feature
Span plan-execModel ClaudeLatency 38mCost $0.00

Heartbeat detection to discover silent breaks

GlassFlow expects to hear from running agents on a regular interval. A frozen agent — one that stopped making progress without crashing — is flagged immediately. In every other tool it looks identical to a slow healthy one.

last heartbeat · just now

Metrics that update while traces are still open

Cost, latency, and token consumption update continuously as the agent runs. In existing tools these are only calculated on closed traces — if 90% of your agents are running, your dashboards reflect the minority that finished.

Total traces
39+6
vs last period
Avg duration
1h 54m
p90 4h 32m

Sessions that never close

For always-on agents there is no trace end. GlassFlow’s session model handles open-ended runs, updating metrics continuously and keeping runs queryable at any point — not just after they finish.

session · agent_ops_007now running
72huptime
12,840total events
3sub-agents
Output quality score