Agent observability
Replay what your agents actually did
Replay of 14 coding-agent runs: tool latency, token spend, policy gates and the approvals that stopped a patch.
Demo · simulated data
14Runs
80%Success rate
$0.72Spend
346.8kTokens
38.4sMean run
Implement database migration to add a new 'status' column to the 'transactions' table
acme/payments-api · agent/implement-database-migration-to-ad
0ms1/17 events · $0.06 · 17.3k tok48.3s
agent
tools
gates
repo
planner · plan- 0msPLANPlanned 3 file changes4.1s
migrations/20231014_add_status_to_transactions.sql
+1 −0
+ALTER TABLE transactions ADD COLUMN status VARCHAR(20) DEFAULT 'pending';ALTER TABLE transactions ADD COLUMN created_at TIMESTAMP;
Tool latency
across all 14 runs
| tool | calls | p50 | p95 | fails |
|---|---|---|---|---|
| shell.exec | 24 | 1.3s | 2.9s | 0 |
| http.fetch | 19 | 587ms | 2.8s | 0 |
| db.query | 13 | 799ms | 2.6s | 0 |
| repo.search | 20 | 303ms | 432ms | 0 |
| fs.write | 37 | 38ms | 148ms | 0 |
| fs.read | 26 | 26ms | 42ms | 0 |
Daily budget
$0.72 of $5.00 spent across 346.8k tokens. Within budget.
14%
Build something like this
Want this kind of visibility over your own agents?
The dashboard is a demo, but the work behind it isn't: tracing agent runs, attributing cost, and putting approval gates where they actually matter. If you're building something similar and want a hand, tell me what you're working on.