Blog/Engineering
Engineering

Microsoft Foundry’s Control Plane: Governing Agents Like Real Services

i
iFynx StudioSep 17, 2026 · 8 min read
microsoft foundry agent control plane

Agents finally get identity cards

Through September 2026, Microsoft has been positioning Azure AI Foundry as the enterprise home for building, deploying, and governing agents — including general availability of Foundry Agent Service observability features and a Foundry Control Plane aimed at fleet-wide oversight. Coverage around NVIDIA GTC 2026 and enterprise explainers such as CryptoBriefing’s Foundry overview converge on the same idea: every agent gets an Entra ID identity, registration into Microsoft Agent 365, publishing paths into Microsoft 365 Copilot and Teams, and policy hooks for which models may be deployed.

Why MENA engineering orgs should pay attention

Banks, telcos, and public-sector vendors in the region already live inside Microsoft estates. Foundry’s bet is that agents without identity, audit, and lease-based recovery will never clear compliance. Long-running hosted agents gaining durable identities and recovery leases (previewed on Microsoft’s 2026 roadmap) matter for workflows that span hours — KYC reviews, claims triage, Arabic document packs — not just chat turns.

Engineering checklist

1. One identity per agent, not per chat thread. Map agents to Entra principals the way you map microservices to service accounts. Rotate secrets; do not let a shared “bot user” become a god mode.

2. Publish where work already happens. Shipping into Teams/Copilot reduces shadow IT — and increases blast radius. Pair publish rights with change windows and rollback owners.

3. Model deployment policies are product constraints. When IT can block unapproved models, product must maintain a fallback model card and UX for “assistant temporarily limited.”

4. Observability is a UX input. Control-plane traces should feed support tooling: which tool failed, which policy blocked an act, what the user saw in Arabic.

5. Physical AI adjacency. Microsoft’s GTC narrative also ties Foundry to NVIDIA Physical AI blueprints. Even if you are not wiring robots, the governance pattern — sense, simulate, act, audit — transfers to any agent that touches the real world.

iFynx takeaway

Foundry’s control plane is the enterprise admission that agents are services. Give them identities, policies, and recovery leases — or watch shadow agents proliferate in personal Copilot sidecars.


Acceptance checklist for an agent control plane

Governing agents like real services means identity, policy, observability, and kill switches — not a spreadsheet of prompts.

  • Every agent has a durable identity, owner, environment, and allowed tool set recorded in the control plane
  • Policy attachments (data classes, regions, spend caps) are evaluated before tool invocation, not after
  • Traces link user/session → agent → tool → downstream system with retention that matches audit needs
  • Human approval gates exist for high-risk actions; the UI shows what will run, in AR and EN
  • Versioned agent configs are roll-forward/roll-back capable without redeploying the whole platform
  • A global kill switch can freeze a class of agents within minutes

If you cannot answer who ran which tool against which customer record last Tuesday, the control plane is theater.


Pitfalls that make “agent platforms” ungovernable

  1. Shadow agents run from laptops with personal API keys outside the control plane
  2. Policies are documented in Confluence but not enforced at the tool gateway
  3. Cost alerts fire weekly while agents burn tokens hourly
  4. Multi-market teams share one agent profile that ignores KSA vs UAE data rules
  5. Incident response has no agent-specific runbook — only model vendor status pages

Acceptance should include a tabletop: revoke an agent mid-flight and prove in-flight calls fail closed.


Ops playbook: four weeks to service-grade agents

  • Week 1: inventory agents; assign owners; disable orphans
  • Week 2: enforce identity + tool allowlists at the gateway for production traffic
  • Week 3: ship bilingual approval UX for the three highest-risk tool classes
  • Week 4: kill-switch drill + audit export sample for compliance
  • Success metric: 100% of production agent invocations attributed to a control-plane identity

Agents become boring infrastructure only when governance is productized.


MENA enterprise notes

Regulated buyers will ask how agents map to existing IAM, DLP, and Arabic customer journeys.

  • Map agent roles to bank/fintech job codes before the first pilot expands
  • Keep approval screens RTL-complete; partial Arabic is a trust defect
  • Document residency of traces and prompt logs per market
  • Align spend caps with procurement — agents are a new cost center, not a free side effect of AI

Identity cards for agents are the minimum viable trust.

Related reading

Treating agents like microservices: what that means in practice

Engineering teams already know how to run services safely: each has an identity, an owner, deployment pipelines, monitoring, alerts and on-call. Agents need the same discipline, plus a few extras because their behaviour is probabilistic.

  • Identity: a dedicated workload identity per agent, with credentials managed and rotated by the platform.
  • Environments: separate development, staging and production agents, with promotion through a pipeline rather than edits in production.
  • Versioning: prompts, tools, model versions and policies versioned together so you can roll back a release as a whole.
  • Observability: traces for each run showing model calls, retrieved documents, tool calls, latency and cost.
  • Service levels: defined targets for latency, success rate and escalation to humans.
  • On-call: someone responsible for responding when an agent misbehaves, with a runbook and a kill switch.

Evaluation as part of the release pipeline

Unit tests do not tell you whether an agent answers correctly. Build an evaluation set for each agent from real (anonymised) requests, including Arabic and mixed-language inputs for MENA users, and run it automatically whenever prompts, tools or model versions change. Track groundedness, task success, refusal correctness and safety. Block releases that drop below agreed thresholds, as you would for failing tests. The NIST AI Risk Management Framework (AI RMF 1.0, January 2023) offers a structure for these activities across govern, map, measure and manage.

Security controls that belong in the platform

The OWASP Top 10 for LLM Applications 2025 lists prompt injection first and expands on excessive agency. Platform-level mitigations include least-privilege tool permissions per agent, content filtering on inputs and outputs, isolation between tenants and between agents, network egress controls and approval steps for high-impact actions. Centralising these in a control plane avoids each team reimplementing them inconsistently.

Data residency and regulated workloads in MENA

Regional banks, telcos and government vendors must know where models run and where prompts, outputs and logs are stored. Choose regions deliberately, document data flows per agent and align them with applicable law, such as the UAE’s federal Personal Data Protection Law (Federal Decree-Law No. 45 of 2021) and Saudi Arabia’s PDPL, as well as sector rules. Long-running agents that process documents for hours, such as KYC reviews, need durable state that stays within the same boundaries.

Cost management

Agents can become expensive quietly, through long contexts, repeated tool calls or retries. Track cost per run and per successful task, set budgets per agent, cache results where safe, and alert on unusual spikes. Cost visibility should sit next to quality metrics so teams can make informed trade-offs.

Platform team responsibilities

As agent numbers grow, a small platform team should own the shared pieces: identity integration, model gateway, policy engine, evaluation tooling, observability dashboards and templates for new agents. Product teams then focus on the agent’s purpose and quality. This mirrors how many organisations run internal developer platforms for microservices, and it prevents every team from solving security and monitoring differently.

Incident response for agents

Define what counts as an agent incident, such as a wrong action, a data exposure or a cost spike, and how to respond: suspend the agent, preserve traces, assess impact, notify affected users where required and fix before re-enabling. Run a drill so the team knows where traces live and who can pull the kill switch.

Common mistakes

  • Running every agent under one shared service account.
  • Editing prompts directly in production without versioning.
  • Monitoring uptime but not answer quality.
  • No owner or on-call for agents after launch.
  • Ignoring cost until the monthly bill arrives.

First 30 days

List every agent in production, assign owners, give each its own identity, add tracing and build a small evaluation set for the most important one. That baseline makes every later platform decision easier and shows quickly where the biggest risks are.

Organisations that treat agents as first-class services from day one avoid the expensive clean-up that follows when dozens of unmanaged agents have already reached production.

Frequently asked questions

What is an agent control plane?

A central layer that registers agents, manages their identities and permissions, applies policies and provides monitoring and audit across the fleet.

Do we need a vendor platform to govern agents?

Not necessarily, but a platform can provide identity, observability and policy features that would be costly to build and maintain yourself.

How do we test an agent before release?

With an evaluation set of realistic tasks, run automatically on every change, with release gates based on agreed quality thresholds.

Who should be on call for agents?

The team that owns the agent, supported by platform engineers for infrastructure issues.

Related: Agent 365 in UAE data centres · Plugin4Shell and coding-agent supply chains. Building enterprise AI agents in the Gulf? See AI development for the UAE and AI development for Saudi Arabia.

For a broader view, see our guide to AI agents for business.