Agents
Registered agents, the version currently serving traffic, and what is staged behind it. When a metric moves, the question is always which of the three inputs moved with it.
Registered
3
Staged changes
2
Runs 24h
321
Cost 24h
$21.60
| Agent | Deployed | Model | Runs 24h | Success | p95 | Cost 24h |
|---|---|---|---|---|---|---|
DataAgent platform-ml ocr.extract_textstorage.write_reportstorage.list_documentsschema.validatenotify.slackocr.health_checkdb.upsert_invoices | v4.2.0 ↑ v4.3.0-rc1 | Sonnet 5 | 35 | 99.4% | 38.8s | $2.85 |
FinanceBot finance-eng pdf.parsedocs.writedocs.fetchmarket.compare_peersnotify.email | v2.8.1 | Opus 5 | 111 | 99.4% | 12.4s | $18.37 |
OpsAgent observability probe.http_checkmetrics.query_latencyledger.fetch_batchledger.match_entries | v1.9.4 ↑ v2.0.0-rc3 | Haiku 4.5 | 175 | 97.3% | 2m 0s | $0.38 |
Staged, not yet deployed
Everything that differs between what is serving traffic and what is waiting behind it. A pass-rate drop next to an empty diff is an upstream problem; a pass-rate drop next to a prompt change is not.
DataAgentv4.2.0 → v4.3.0-rc1
- promptAdded explicit instruction to retry OCR once before falling back to pdf.parse.
- toolsGrant added: pdf.parse. Previously reached only through the OCR fallback path.
OpsAgentv1.9.4 → v2.0.0-rc3
- modelHaiku 4.5 → Sonnet 5. Health-check summaries were losing detail under load.
- promptEscalation threshold moved from 3 consecutive failures to 2.