Service Desk AgentMicrosoft Foundry · production
- Owner
- Service Operations
- Quality
- 92% · passed
- Authority
- Read + prepare
Why this service exists
An agent can be available, useful, and still be ungoverned.
Traditional monitoring asks whether an application is running. Agent governance must also ask whether it should run, for whom, against which evidence, using what identity, with which tools, under whose authority, at what cost, and until when.
Agent AI governance is the operating discipline that keeps those answers true after launch.Ownership changes
A maker leaves, a business owner changes roles, or the original sponsor disappears while the agent, credentials, audience, and scheduled actions remain active.
Authority expands
A new connector, application permission, maker-provided credential, sharing change, or anytime operation gives the agent more reach than its original approval covered.
Evidence drifts
A SharePoint library moves, a policy changes, a source owner stops maintaining content, or web grounding introduces information outside the approved evidence boundary.
Behavior changes
Models, prompts, orchestration, tools, dependencies, and safety configurations change independently, so yesterday’s evaluation cannot prove today’s result.
Failures hide in success
The endpoint remains available while groundedness, completion, tool choice, user trust, latency, retries, quota, or cost quietly moves outside the operating envelope.
Agents outlive their purpose
Pilots become permanent, duplicates accumulate, review dates expire, production and development blur, and no one makes an explicit renewal, exception, or retirement decision.
Source moved→Retrieval weakened→Answers stayed plausible→Quality incident
Maker left→Credential remained→Actions continued→Authority incident
Audience widened→Data boundary shifted→Approval no longer matched→Governance incident
The governance operating system
A registry becomes useful when every change creates a decision.
Temen begins with read-only inventory, correlates technical evidence with business decisions, applies a governance model, and then operates the reviews, exceptions, releases, incidents, and retirement actions that keep the registry alive.
Know every custom, publisher-managed, draft, published, quarantined, and retired agent across environments and channels.
Record the business purpose, audience, decision impact, data sensitivity, risk tier, prohibited use, and the reason the agent should exist.
Name active business and technical owners, publishing authority, support owner, risk acceptor, and the next review date.
Review authentication, agent identity, linked applications, permissions, groups, tenant-wide sharing, external channels, and consent.
Inspect sources, web search, connectors, operations, credential mode, anytime execution, approvals, reversibility, and logging.
Separate development, test, and production; apply managed-environment controls, policies, evaluation gates, publishing approval, and rollback.
Monitor health, task quality, groundedness, safety, failures, latency, usage, quota, caching, cost, and user feedback together.
Review exceptions, renew approvals, track material change, preserve decisions, test retirement, and close ownership when the service ends.
DISCOVER→REGISTER→RISK-TIER→APPROVE→OBSERVE→CHANGE→ATTEST→RETIRE
Interactive governance demonstration
Break an approved agent. Watch governance respond.
Start with an approved production record, then deliberately remove the owner, widen the audience, use a maker credential, allow autonomous actions, move environments, or expire the review date.
GOVERNANCE POLICY LABExpense Approval Agent · AGT-0047
Approved baselineGovernance score100/ 100
Blocked controls0must resolve
Review controls0owner decision
LifecyclePRODapproved v2.4
Break the approved baseline.
Each action simulates a real configuration or lifecycle change. Watch the registry, score, findings, and release decision respond.
REGISTRY RECORD
- Business purpose
- Prepare expense decisions
- Risk tier
- Moderate
- Business owner
- Finance Operations
- Next review
- 2026-10-10
Accountable owner
Business + technical owner active
Attestation currentAuthentication
Microsoft Entra required
Identity boundary verifiedAudience
Pilot security group
Audience matches approvalTool credential
End-user identity
User context retainedAction authority
Approval before write
Human checkpoint retainedEnvironment
Managed production
Policies and separation verifiedLifecycle review
Quarterly review current
Next review scheduledWhat Temen operates
Governance and support work as one service.
The managed-service offer covers ongoing maintenance, operational support, visibility, controlled optimization, and reporting. The governance layer gives every action a purpose, owner, boundary, decision, and retained record.
Ongoing maintenance
Minor prompt revisions, tuning, controlled improvements, workflow corrections, source maintenance, dependency updates, and approved configuration care.
Operational visibility
Agent runs, failures, quality, usage, quotas, caching, model behavior, connectors, actions, latency, and successful-task cost.
Incident response
A defined path to triage, contain, restore, preserve evidence, communicate impact, verify recovery, and address recurring causes.
Controlled optimization
Model and prompt selection, routing, caching, context, tool efficiency, performance, reliability, and cost based on observed evidence.
Managed governance
Registry, owner attestations, risk-tier reviews, approved environments, connector and credential standards, publishing gates, exceptions, and retirement.
Service reporting
Health and change summaries, recommendations from usage and telemetry, governance findings, risk decisions, capacity signals, and improvement priorities.
Interactive operations demonstration
A running agent still needs quality, cost, and access decisions.
Select a staged production signal and move through containment, investigation, regression, recovery, and evidence. Uptime is shown beside the signal so the distinction is impossible to miss.
SERVICE MONITORINGAgent Operations · staged tenant
Last evaluation 08:42 UTCAvailability99.93%within targetQuality gate71%below 90%Task cost$0.084watchOpen changes021 awaiting approval
CONTROLLED RESPONSEINC-0042
Grounded-answer score fell after source update
- 1
Reduce exposure to affected source
- 2
Compare source and retrieval changes
- 3
Run regression evaluation
- 4
Restore known-safe configuration
Interactive change-control demonstration
Every request enters through the same door. It does not leave through the same lane.
Compare a maintenance repair, a controlled model change, a high-risk connector expansion, and a net-new external agent. Advance each request through the evidence and approval path it actually requires.
CHANGE CONTROLAgent service request queue
STAGED WORKBENCHREQUESTCR-0118
Repair a stale source
Knowledge Assistant stopped citing the current support runbook after a SharePoint library move.
- Service lane
- Managed maintenance
- Risk
- Low
- Operating decision
- Restore the approved evidence path without changing the agent’s business outcome or authority.
REQUIRED EVIDENCE
- Affected source and owner
- Known-good answer set
- Citation comparison
- Version and rollback record
CONTROL PATH0 / 6
- 01
Request recorded
- 02
Impact classified
- 03
Source path tested
- 04
Regression set passed
- 05
Approved version released
- 06
24-hour observation closed
What managed governance looks like
Four examples where “the agent is up” would be the wrong conclusion.
These staged cases show how Temen joins runtime evidence to the agent registry, operating contract, owner decisions, change record, and recovery proof.
The agent stayed online after its authoritative source changed.
Signal
Grounded-answer pass rate fell from 94% to 71% after a support library restructure.
Governance context
The source owner, approved corpus, quality threshold, and rollback version were already recorded.
Temen response
Temen reduced exposure to the affected source, compared retrieval changes, ran the regression set, restored the known-safe configuration, and retained the incident evidence.
A maker left while an agent still held a shared write connection.
Signal
The technical owner account became disabled, but a maker-provided credential and anytime operation remained attached.
Governance context
The registry correlated ownership, credential mode, connector operations, audience, linked identity, and lifecycle status.
Temen response
Temen blocked publishing and write actions, assigned an interim owner, reviewed delegated authority, replaced the connection design, tested recovery, and scheduled attestation.
Retries made the workload look popular while successful-task cost climbed.
Signal
Token consumption rose 38%, but the change came from long context, repeated tool calls, and failed retries rather than new business value.
Governance context
The workload had an approved cost envelope, model and routing record, retry policy, cache policy, and owner for economic decisions.
Temen response
Temen separated useful runs from retries, inspected traces, tightened context and tool calls, applied safe caching and routing, and confirmed the revised service baseline.
A useful internal assistant was proposed for a public transactional role.
Signal
The request added an external audience, customer data, eligibility recommendations, new APIs, write actions, and a different incident profile.
Governance context
The change gate compared the request with the approved purpose, audience, data, authority, architecture, risk tier, and service boundary.
Temen response
Temen classified the request as a separate project, preserved the existing production service, and opened discovery for privacy, threat modeling, architecture, evaluation, pilot, and support.
From assessment to managed rhythm
The first 90 days establish control. The service keeps it current.
Temen’s read-only assessment creates the initial technical registry and governance findings. Business owners then supply the purpose, risk, approval, review, lifecycle, and exception decisions that configuration cannot infer.
Confirm inventory coverage, organization-built and publisher-managed scope, owners, purposes, risk tiers, approved environments, audiences, identities, connectors, credentials, actions, and immediate exceptions.
Define publishing gates, owner attestations, environment and connector standards, identity and credential patterns, evaluation requirements, exception handling, incident ownership, versioning, and rollback.
Complete the first attestation, resolve high-risk findings, test quarantine or retirement, publish the service baseline, rehearse incident and change paths, and establish reporting.
Published starting packages
Match cadence and oversight to business impact.
Final scope and pricing follow validation of workload count, architecture, business impact, telemetry, usage, dependencies, service hours, and complexity. Model, token, search, storage, voice, connector, licensing, and third-party charges remain separate unless stated otherwise.
Agent Care
$495/month1 standard AI workloadReactive support, monthly health review, minor prompt revisions, basic workflow corrections, two engineering hours, and a brief monthly summary.
- Defined support path
- Monthly service health
- Minor approved maintenance
Agent Operations
$995/monthUp to 3 standard workloadsProactive monitoring, prompt tuning, workflow maintenance, biweekly health review, five engineering hours, and monthly recommendations.
- Proactive operational visibility
- Biweekly review and tuning
- Monthly recommendations
Agent Assurance
From $1,995/monthUp to 5 adjusted workloadsEnhanced oversight, active tuning, workflow optimization, weekly performance review, ten engineering hours, and executive reporting.
- Enhanced governance oversight
- Weekly performance review
- Executive and technical reporting
Customer operating evidence
Governance is visible in the records people use to decide.
The service does not end with a dashboard. Temen maintains the registry, business decisions, service baseline, release evidence, incident history, and reporting needed for leaders, administrators, makers, security, support, and risk owners to act.
Tenant agent registry
Agent identity, platform, state, owner, environment, audience, authentication, channels, knowledge, connectors, operations, and linked identities.
Business governance overlay
Purpose, business owner, risk tier, approval status, review dates, lifecycle state, exceptions, and decision notes that configuration alone cannot prove.
Operating baseline
Approved versions, quality and safety thresholds, availability, latency, usage, quota, cache, cost envelope, dependencies, and known limitations.
Policy and release evidence
Environment, identity, sharing, credential, connector, publishing, test, approval, version, rollback, and observation records.
Incident and change records
Signals, impact, containment, actions, communications, recovery proof, root causes, requests, approvals, releases, and follow-up decisions.
Service and governance report
Health, quality, cost, changes, exceptions, owner attestations, review status, risks, recommendations, capacity, and improvement backlog.