Observability & governance
If agent traffic crosses the gateway, you can trace it.
For model calls, tool invocations, and agent turns routed through AgentGuide, observability is a property of the traffic rather than an SDK each team must integrate. The same position makes supported gateway controls enforceable, not advisory.
Usage events
Four event families, one attribution model.
LLM, MCP, ACP, and builtin traffic each produce persisted, typed usage events. Unified AgentRoute ingress stamps the target agent id and selected runtime directly for both ACP and builtin turns. Direct LLM and MCP traffic remains visible without inventing an owner.
| time | kind | agent | target | tokens in / out | latency |
|---|---|---|---|---|---|
| 14:02:11 | llm | support-triage | anthropic · claude-sonnet-5 | 8,412 / 1,203 | 2.1s |
| 14:02:13 | mcp | support-triage | crm-tools · lookup_account | — | 340ms |
| 14:02:19 | builtin | report-writer | turn · plan-execute | 31,908 / 4,551 | 18.7s |
| 14:02:24 | acp | codex-dev | turn · session 7f3a | 12,050 / 2,340 | timeout |
Cost-grade token detail
Prompt, completion, cached, and reasoning tokens are recorded separately — the numbers you actually need to explain a bill or tune a cache.
Spans, not just counters
Agent turns open interaction spans; inner model calls and tool executions nest underneath. A stable run id connects builtin HITL segments, while resumed and expired segments link to the checkpoint-producing trace in OTLP.
Failures classified
Client errors, upstream failures, and internal faults are distinguished per event — so an agent misbehaving and a provider outage never blur together.
Analytics & integration
Query it here. Aggregate it anywhere.
Metrics Admin API
Summaries with pipeline health, per-protocol timeseries and breakdowns, recent interaction events, and per-agent usage and activity views — queryable the moment the gateway is running.
- Per-agent rollups Usage, activity, interactions, and health per registered agent.
- Prometheus exposition Scrape the gateway; build trends and alerting in Grafana.
- OpenTelemetry export Send reconstructed interaction spans over OTLP gRPC or HTTP, with optional eino component spans.
- Retention control Configurable retention with automatic cleanup — observability that respects your storage.
Governance
Enforcement at the only place it can't be skipped.
Guidelines in a wiki don't stop an agent at 3 a.m. Controls at the gateway do. AgentGuide's governance is structural: if traffic doesn't satisfy policy, it doesn't reach the model, the tool, or the agent.
Identity & keys
- Virtual keys per app, team, or agent — issue, rotate, revoke centrally
- Upstream secrets never leave the gateway
- Route-level auth requirements, fail-closed
Permission gates
- Deny by default; auto-approve where safe; interactive human approval where it counts
- Unanswered requests time out to denial
- Permission requests, resumptions, cancellations, and expiries share correlation ids; a dedicated compliance audit ledger remains roadmap work
Runaway protection
- Agent recursion-depth limits stop agent-calls-agent loops at the gate
- Per-agent concurrency caps and turn timeouts reject rather than queue
- Panic containment: one bad turn never takes down the gateway
Scoped capability
- Builtin definitions enforce their selected model route and declared MCP tools; generic resource links are a management view
- ACP sessions may start only under configured allowed roots; OS-level sandboxing remains the deployer's responsibility
- Missing tools fail loudly — no silent fallback behavior
- One route, one owning agent: attribution stays unambiguous
Put your agents on the radar
Start observing before you scale to a fleet.
The teams that succeed with multi-agent systems are the ones that could see their single agents first. Deploy the gateway, route one agent through it, and watch the picture appear.