The 200 problem
APM assumes a deterministic code path. An agent that chooses the wrong tool completes successfully, so every threshold stays green while the answer is wrong. The failure is a judgement, and judgements do not raise exceptions.
SERAA Cortex
Agent Observability
for Enterprise AI
One record for what an agent did, what it decided, and what it cost.
An agent that picks the wrong tool returns HTTP 200 and a confident wrong answer. Nothing in an APM suite catches that, because nothing failed. The Control Tower instruments the decision rather than the request, and puts the trace of a run and the bill for that run on the same record.
Every callLogged with per-call cost
Every stepTool calls, latency, errors
Every agentBudget ceiling and kill switch
StandaloneBuy it without the suite
01 · Definition
What agent observability actually is.
Agent observability is the practice of recording what an AI agent did, why it did it, what it cost and whether it stayed inside policy, in enough detail to reconstruct any single run afterwards.
It differs from application monitoring in what counts as a failure. A web service fails by throwing an error. An agent fails by succeeding at the wrong thing: it picks the wrong tool, retrieves the wrong document, or answers fluently from nothing, and returns a 200 while doing it. So the unit of observation is not the request. It is the decision.
That distinction decides everything downstream. If you instrument requests you get latency and token counts, which tell you an agent was busy. If you instrument decisions you get the evidence chain behind an answer, which is the only thing that lets somebody accountable for the outcome sign off on it.
The short version
Four signals make a run reconstructible: the trace of the steps, the log of the calls, the live status of the agent, and the controls you can pull while it is running. Miss any one of them and you have a dashboard rather than a governance plane.
The four signals
TraceEvery step, tool call, error
LogEvery LLM call, with cost
StatusLive, per agent
ControlGuardrails, budgets, kill switch
CorrelationSession ID across a run
RetentionQueryable audit trail
02 · The gap
Why the tools you already pay for do not cover this.
Most agent estates fail their first audit and their first budget review at the same time, for the same reason. The traces live in one vendor’s console, the invoices in another, and nothing reconciles the two.
The bill is a separate system
Token spend arrives on a provider invoice weeks later, aggregated past the point of attribution. By the time finance asks which agent caused the jump, the traces that would answer it have rolled off.
Volume pricing punishes agents
Tools metered on ingested spans were priced for single model calls. A multi-step agent run emits eight to fifteen. The bill scales with how hard the agent thought, which is exactly the wrong incentive to put on an engineering team.
The operator is not the author
Developer tracing tools are built for the engineer who wrote the agent. The person who has to run it on a Tuesday, and stop it, is somebody else, and usually cannot read a span waterfall.
03 · The surface
What the Control Tower records, and what it can do about it.
Recorded on every run
A trace of each step the agent took, including tool and MCP calls, latency and errors. A log line for every LLM API call carrying timestamp, status, session ID, agent name, model, provider, token counts and per-call cost. The session ID stitches the calls in one run together, so a bad outcome is followed back to the step that caused it rather than guessed at.
Attributed to something a business recognises
Cost is tracked per request and attributed to agent, model, provider and the user who triggered the run, with ranked views of the highest-cost agents, models, users and providers. A business unit can be shown what its own agents cost instead of a share of one undifferentiated invoice.
Watched without anyone watching
Live operational status for every deployed agent with success rates, provider detail and Active or Inactive state. Guardrails enforce content filters, PII detection and redaction, output-length limits and compliance policy including GDPR and HIPAA. A degrading agent surfaces as a trend, not as a support ticket.
Stoppable
Budget thresholds are set platform-wide or per agent, in advance, to stop overruns rather than report them. Any agent below its defined performance, accuracy or bias thresholds can be suspended immediately. That is what makes deploying an agent a reversible decision.
Deeper reference
The full capability table for the Control Tower module, and the same for Agent Work Bench, Multi-Agent Orchestration, Agent Registry and System Config, sits on the SERAA Cortex page, alongside the control plane diagram and the FinOps detail at Inside the Control Tower.
Observability
LLM call logEvery call, searchable
Per callTokens, status, cost
CorrelationSession ID
Trace detailTool calls, latency, errors
Activity logReal-time audit trail
Control
GuardrailsContent, PII, length, policy
Budget ceilingsPlatform-wide, per agent
Kill switchPer agent
AccessSBAC, RBAC, ABAC
RuntimesGKE, AKS, EKS, Vertex
04 · The landscape
Four categories of tool, sold to four different people.
Agent observability is bought from three directions at once, and the categories are not substitutes for each other. This is how each is generally sold, so you can tell which conversation you are actually in.
| SERAA Agent Control Tower |
Governance control towers ServiceNow, IBM watsonx Orchestrate |
Observability suites Datadog, Dynatrace, New Relic |
Developer eval and tracing LangSmith, Langfuse, Arize, Braintrust |
|
|---|---|---|---|---|
| Who it is built for | Whoever is accountable for agents in production | Risk, compliance and the CIO | The platform team that already owns the estate | The engineer who wrote the agent |
| Unit you are billed on | Governed agents and agent steps | Platform subscription, quoted | Ingested spans or events, plus hosts | Traces, spans, scores or seats |
| Entry price published | Yes, indicative list below | No, sales contact only | Yes, with volume overage | Yes, most publish a low tier |
| Trace and per-call cost on one record | Yes, by default | Varies by module | Cost is usually a separate product | Usually token counting rather than spend |
| Runtime control, not only visibility | Kill switch and budget ceilings per agent | Policy and approval workflow | Alerting | Rarely, these are read paths |
| Agents built somewhere else | Cost and usage, when they route through the Tower as AI gateway | Broad discovery is the pitch | Whatever emits telemetry into the suite | Whatever the SDK is wired into |
| Inside your own boundary | SaaS or sovereign VPC. Air-gapped at design-partner stage | Mostly vendor cloud | Mostly vendor cloud | Several offer self-hosting |
| Take the module on its own | Yes, the Control Tower ships standalone | Usually platform-wide adoption | Bundled with the suite | Yes |
Columns describe how each category is generally sold, based on what those vendors publish on their own pages as of August 2026. It is a comparison of categories, not a set of claims about individual products, and we will correct anything a vendor tells us is wrong. Prices and packaging change; check the source before relying on any of it.
05 · Pricing
Priced on governed agents, not on how hard they thought.
Per-seat pricing does not survive contact with agents, and span-volume pricing makes the bill a function of reasoning depth. Cortex is priced on governed agents and agent steps, which is the axis a business can forecast.
SaaS
From $8K per month
Business tier, annual and sales-assisted, with multiple scopes and unlimited projects, SSO, audit export and delegated admin, committed usage with overage. Before that there is a two-week free trial that needs no card and no sales call. Enterprise is custom, and covers dedicated or sovereign tenancy, ISO 42001, GDPR, HIPAA and EU AI Act obligations, and a contracted SLA.
Prices indicative
Sovereign
From $10K per month
An annual platform licence plus a declared governed-agent band, so no usage telemetry has to leave your boundary. The band is self-attested and reconciled at renewal rather than mid-quarter. Carries a 15 to 25 per cent premium over the equivalent SaaS list price, which is deployment and support cost rather than scarcity.
Your VPC or data centre
Pass-through
No margin on inference
Compute and model consumption run on your own accounts and provider invoices at actuals, auditable against your own cloud bill. We make margin on the platform, not on reselling your tokens back to you. It also means the Tower’s cost attribution can be reconciled against a bill you already receive.
Actuals, auditable
Why the unit matters
A single model call emits one span. One multi-step agent run emits eight to fifteen. Anything metered on ingested volume therefore bills you more for an agent that plans properly than for one that guesses, and nobody can forecast next quarter from it. Governed agents and agent steps are countable in advance, which is the whole point.
06 · The condition
One architectural decision, worth taking early.
Agents built on third-party platforms, including Gemini Enterprise, Azure AI Foundry, AWS Bedrock and Databricks AgentBricks, are monitored for cost and usage provided they route through the Control Tower as their AI gateway.
That is the one dependency this rests on, and it is better met in the first week than discovered in the first audit. Routing through the gateway is what makes one log the system of record instead of four partial ones. We put it on the page rather than in an implementation note because a buyer should price the decision in, not around it.
What holds either way
Vendor neutralGKE, AKS, EKS, Vertex
Standalone moduleNo suite required
Agents already builtStay where they are
Air-gappedDesign-partner stage
07 · Questions
The questions that come up in every first conversation.
Agent observability is the practice of recording what an AI agent did, why it did it, what it cost and whether it stayed inside policy, in enough detail to reconstruct any single run afterwards. The unit of observation is the decision the agent made, not the request it served.
Application monitoring assumes a deterministic code path, so a failure throws an error. An agent fails by succeeding at the wrong thing: it picks the wrong tool, retrieves the wrong document, or answers confidently from nothing, and returns HTTP 200 while doing it. LLM observability is closer but usually stops at the model call. Agent observability has to cover the steps between the calls, which is where most agent failures actually live.
A searchable log of every LLM API call with timestamp, status, session ID, agent name, model, provider, token counts and per-call cost. A trace of every step an agent took to produce a response, including tool calls, latency and errors. Live operational status for every deployed agent with success rates and Active or Inactive state. A real-time activity audit log. Guardrail and policy enforcement across content filters, PII detection and redaction, output-length limits and compliance policies including GDPR and HIPAA.
Yes, for cost and usage, provided those agents route through the Control Tower as their AI gateway. That covers agents built on Gemini Enterprise, Azure AI Foundry, AWS Bedrock and Databricks AgentBricks. It is one architectural decision and it is better taken deliberately at the start than discovered later, so we raise it in the first session rather than the fifth.
Both, and the second is the part that matters to whoever signs off the deployment. Any agent that falls below its defined performance, accuracy or bias thresholds can be suspended immediately with a per-agent kill switch. Budget ceilings are set platform-wide or per agent to stop cost overruns rather than report them. Visibility without a control is a dashboard, not a governance plane.
Cortex is priced on governed agents and agent steps, because per-seat pricing does not survive contact with agents and span-volume pricing makes the bill impossible to forecast on a multi-step workload. There is a two-week free trial that needs no card, a sales-assisted Business tier from 8,000 US dollars a month, and a custom Enterprise tier. Sovereign deployment starts at 10,000 US dollars a month and carries a 15 to 25 per cent premium over the equivalent SaaS list price. Figures are indicative.
No. Infrastructure and model consumption run on your own accounts and provider invoices at actuals, auditable against your own cloud bill. Covasant makes margin on the platform, not on reselling your tokens back to you.
Yes. The Control Tower runs as SaaS or inside your own VPC, and is vendor-neutral across GKE, AKS and EKS plus GCP Vertex AI Agent Engine. Fully air-gapped deployment is at design-partner stage, which is a different claim from sovereign VPC and we keep the two separate on purpose.
Yes. It ships as an independent offering, as do Agent Work Bench and Agent Registry. You can monitor and control production agents without adopting the rest of the suite and without moving the agents you already have.
The evidence an auditor asks for is the log, the trace and the access record, and all three are queryable rather than reconstructed. Access is governed through Scope-Based Access Control across System, Organization and Project levels, with role-based and attribute-based controls on top. Enterprise agreements cover ISO 42001, GDPR, HIPAA and EU AI Act obligations.
Related components
The Control Tower sits inside Cortex. Cortex is one of four.
Get started
Bring one agent estate nobody can currently see.
We will show you what governing it looks like: the log, the trace, the cost attribution and the kill switch, running on your own cloud.