Loading…
You cannot govern a system you cannot see. We instrument the systems we build to emit structured telemetry from every surface — agent messages, policy-gate decisions, plan-step completions, health signals — and make it queryable in real time. Not after the fact. Not by reading logs. In real time, with the ability for your team to steer or pause any agent or plan from the same interface.
Observability isn't a separate layer you bolt on later — we build it into the system from the start. Every surface emits structured telemetry automatically. The observability layer aggregates, indexes, and exposes it — from every agent, in real time.
Read the observability docs ↗A2A exchanges, policy-gate decisions, plan-step completions, tool calls, and security events are all structured telemetry. There is no instrumentation code for you to write — we wire the emission into the system itself. Every event carries the agent identity, the surface, the action, the outcome, and a high-resolution timestamp.
The event stream is written to a queryable store with automatic indexing by agent identity, plan ID, event type, and time. Single-agent traces resolve fast enough to query interactively. The retention window is configurable to your policy. Events are immutable after write.
The control plane assembles full traces for any agent or plan: the complete ordered sequence of events, their inputs and outputs, their latencies, and their relationships to upstream and downstream agents. Traces are available in real time — you don't wait for a batch job.
Pattern-matching rules run continuously against the event stream. When a rule fires — a permission violation, an unusual delegation chain, a plan step exceeding its SLA — an alert is emitted as a structured event and routed to your configured receiver. You do not manually tail logs to find problems.
The control plane exposes steering actions for every running agent and plan: pause execution, drain active connections, redirect a plan step to a different agent, or shut down cleanly. Steering actions are themselves logged as events. You never need to touch code or restart a deployment to intervene in a running agent.
Every agent action is a queryable event with full context: agent identity, action, inputs, outputs, outcome, and timestamp. When something goes wrong — or when an auditor asks — the answer is in the trace. You are not reconstructing it from fragmented application logs or asking the agent what it did.
Pattern-matching rules run continuously against the live event stream. A permission violation, an unusual delegation chain, a plan step overrunning its SLA — these appear as alerts in real time, not in the post-incident review. Your team sees the signal when it can still act on it.
Pausing, redirecting, or shutting down a running agent is a control-plane action — not a deployment. The change is instant, logged, and reversible. Your on-call engineer can intervene in a running agent at 2am without needing to touch infrastructure, modify code, or wait for a deployment pipeline.
Telemetry emission is asynchronous and batched — agents do not block on it. The structured events are written to a queryable store off the hot path. In practice the overhead on agent execution is negligible; the cost is in the telemetry store, which is sized for your event volume and retention window, not in agent latency.
Yes. The systems we build emit OpenTelemetry-compatible structured events, so they map into existing pipelines — Grafana, Datadog, an in-house SIEM. You can run the control surface we ship and forward the same event stream to your established tooling; they are not mutually exclusive.
Pause or resume any agent, redirect or cancel a running plan, revoke a delegation, and shut down a misbehaving agent — all from the same surface that shows you the telemetry. Detection and control share one plane, so the path from a signal to a corrective action is as short as possible.
As far back as your retention policy declares. The audit log is append-only and the telemetry store keeps events for the window you configure. For regulated workloads that require multi-year retention, the store is sized accordingly; for high-volume low-retention cases, you can keep a shorter window with materialized rollups for the long tail.
Yes. The telemetry pipeline, the query store, and the control surface all run on-prem with no cloud dependency, the same as the rest of the system. Air-gapped deployments get the full real-time fleet view with no data leaving the environment.
Tell us what you're building. A real engineer replies.