Billing day: Cowork and Work IQ go GA, and the agent meter starts running
June 16 was billing day. Copilot Cowork went generally available worldwide, a month early, and the Work IQ APIs reached GA the same day. Both meter in Copilot Credits. The tooling to run and watch metered agents matured around them. And the layers you do not own kept moving: a frontier model stayed export-suspended, five Llama models retired in Foundry, and the EU postponed the AI Act's high-risk dates by seventeen months.
This week's thread: the consumption economy for agents went live
Build gave agents a meter as a slide. This week it started running. On June 16 Copilot Cowork and the Work IQ APIs both hit GA, both billed in Copilot Credits at a cent each. Agent work is now a consumption line item, not a per-seat one. The apparatus grew around it the same week: @Jared Spataro's Cowork reached Dynamics 365, the Power CAT kit became the Copilot Agent Kit with a credit-consumption analyzer, Sentinel began ingesting agent identities, and the runtime story moved onto Kubernetes. @Charles Lamanna's platform turned the meter on. Underneath, the model and policy layers reminded you the part you bill for is not the part you control: Fable 5 stayed dark, five Llama models retired, and the EU bought everyone seventeen more months.
Signal of the week
Copilot Cowork is generally available, and it is metered
On June 16 Copilot Cowork went generally available worldwide for Microsoft 365 Copilot customers, a month ahead of its July forecast. It plans and works through multi-step tasks across your tools while you review the output. It runs on Anthropic Opus 4.8 and Sonnet 4.6, with GPT-5.5 in Frontier and a cheaper Cowork 1 model coming. Pricing is the headline: usage-based in Copilot Credits at a cent each, charged on model use, context, tool calls, and runtime. Nine partner plugins are live, with audit, DSPM, eDiscovery, and Insider Risk Management at GA.
Why it matters
A Copilot task now costs by consumption, not by seat. A well-scoped one is cheap; a looping one is a cost incident. Frontier tenants are not billed until July 1, so set per-user and per-group credit spend limits this week, and pilot on bounded tasks you can measure.
1The agent meter goes live
The Work IQ APIs reach GA, billed in Copilot Credits
The same June 16 turned on the Work IQ APIs, the grounding surface from Build. They are GA and metered in Copilot Credits: fixed for the Tools domain, variable for Chat and Context. There is no separate SKU. First-party use through Copilot and Cowork is covered by the M365 Copilot license; third-party and custom agents that ground in your tenant pay on consumption, against your credit pool.
Why it matters
Two metered surfaces GA on one day makes June 16 the start of agent FinOps. Stand up the cost-management dashboard and set tenant and group spend limits, so a runaway agent is a capped budget event, not a surprise invoice.
2Copilot Studio and the agent toolchain
The Power CAT kit becomes the Copilot Agent Kit, and it can watch credit burn
The June 2026 Power CAT Copilot Studio Kit release becomes the Copilot Agent Kit, now spanning Copilot Studio, Microsoft 365 Copilot, Agent 365, and Foundry. It adds an Agent Metrics Dashboard, a Conversation Analyzer that reports coverage, sentiment, and credit consumption, Power Shield health views, Wave 2 templates, and CI/CD for the Agent Review Tool. It stays a free, community-supported accelerator.
Why it matters
The credit-consumption view lands the same week consumption became billable. The Conversation Analyzer is the closest thing to a per-agent FinOps cockpit - pair it with spend limits to see which agent is expensive and why, before the tenant-wide invoice.
Agent Framework Python 1.9.0 adds loop control and denies MCP sampling by default
Agent Framework Python 1.9.0, shipped June 18, adds AgentLoopMiddleware for re-running an agent in a loop, and explicit sampling controls. Notably, server-initiated MCP sampling is now denied by default, closing a path where a connected MCP server could make your model generate without sign-off. It also adds Azure AI Content Understanding support and pre-1.0 to_llm_input changes.
Why it matters
The security default is the line that matters: an upgrade quietly tightens what a remote MCP server can make your model do. Read the to_llm_input change before bumping the version in anything shipped.
3First-party agents and Cowork's reach
Planner Agent hits GA, and Cowork extends into Dynamics 365
Two GA moves widened what you can deploy without building. The Planner Agent in Microsoft 365 Copilot hit GA on June 15 - task cards, insights, and natural-language task management, no custom build. A day later Copilot Cowork for Dynamics 365 went GA with Sales and Customer Service plugins that run over MCP, enforce the existing permission model, and keep a human in the loop: Cowork proposes, a person commits.
Why it matters
Planner Agent is the easy adoption win - first-party, deploy now. Cowork for Dynamics is the strategic one: agentic work reached CRM data under the existing security model, so scoping now starts with what Cowork already does here.
4Where agents run, and who watches them
Microsoft Sentinel starts treating agent identities as first-class SOC assets
A public-preview Agent Identities Asset Connector for Microsoft Sentinel, announced June 16, brings agent identity context into the SIEM. It ingests from Agent 365 and Microsoft 365 Copilot, so security teams can hunt over and govern autonomous agents alongside users and devices - the SOC-side complement to the Agent 365 registry.
Why it matters
With agents acting under their own identities and metering credits, 'which agent did what, when, at what cost' is now a security question. Ingesting them into Sentinel is the connector that lets a CISO say yes to an Agent 365 rollout.
The agent runtime story moves onto Kubernetes at Open Source Summit India
At Open Source Summit India and KubeCon India, recapped June 17, the production-runtime side got attention. The AKS updates target agent workloads: agent pool rollback, Fleet Manager enhancements, GPU optimization, faster provisioning, and the Azure SRE Agent for incident diagnosis. Microsoft framed it against its open-source agent stack - the Agent Framework and the open-source Agent Governance Toolkit.
Why it matters
Most agent conversations stop at Copilot Studio or Foundry and never ask where a custom agent runs. Often the answer is AKS, and its tooling is maturing fast. The runtime and its governance belong on the architecture diagram, not as an afterthought.
5Foundry and the learning loop
Foundry publishes a blueprint for owning your reinforcement-learning loop, not renting it
A June 18 Foundry post shows how to build enterprise reinforcement-learning systems that improve through a loop you own, by combining Foundry with the open OpenEnv standard - a small reset, step, and state contract, carried over MCP and packaged with Docker. It assembles Build blocks (Frontier Tuning, Foundry IQ, Memory, Managed Compute) and names McKinsey, Bristol Myers Squibb, and Land O'Lakes as early adopters.
Why it matters
Read against the same week's model news: if a rented model can go dark, the durable asset is the learning loop, not the weights. Most clients are far from parametric RL, but the non-parametric path - optimizing prompts, skills, and tools against outcomes - is reachable now.
6The model and regulation layer
The model layer stayed unstable: Fable 5 still suspended, five Llama models retired
The supply-chain fracture did not heal. Anthropic's Claude Fable 5 and Mythos 5, suspended June 12 under a US export-control directive, stayed offline; Anthropic said it was working to restore access, but the models have not come back. Separately, five Meta Llama models reached retirement in Foundry on June 13 - the 3.1 and 3.2 families. Llama-3.3-70B-Instruct and the Llama-4 models remain GA.
Why it matters
Two mechanisms, one lesson: a model can vanish by government order or by a routine deprecation. The defence is portability - an abstraction layer, fallbacks, and evals. Concretely: migrate any retired Llama agent to Llama-3.3-70B or Llama-4-Scout now.
The EU Parliament postpones the AI Act's high-risk obligations by seventeen months
On June 16 the European Parliament approved the Digital Omnibus on AI, 423 to 57 with 174 abstentions. It keeps the risk-based core but pushes the heavy dates: stand-alone high-risk systems (Annex III) to December 2, 2027, and embedded ones (Annex I) to August 2, 2028. Most other provisions still start August 2, 2026. It also bans AI tools that generate child sexual abuse material or non-consensual intimate imagery.
Why it matters
The breathing room is real but narrow: high-risk obligations slip to late 2027, but the risk-based core and the August 2026 baseline did not move. The governance you build now - Agent 365, Sentinel agent identities, Purview - is still the readiness work.
Voices to follow
- Microsoft EVP - Copilot, Agents & PlatformCharles Lamanna
Executive Vice President for Copilot, Agents, and Platform at Microsoft. He owns the strategy behind Copilot Studio, Agent 365, and the Power Platform, so when the agent stack's direction shifts, it usually shifts because of a decision his org made.
Microsoft EVP - Copilot, Agents & Platform - Microsoft Principal Cloud Advocate - Power Platform + agentsDaniel Laskewitz
Principal Cloud Advocate at Microsoft, ex-Power Platform MVP and co-founder of Forward Forever. One of the clearest voices on multi-agent systems and Power Platform governance - the bridge between the maker community and the product teams.
Microsoft Principal Cloud Advocate - Power Platform governance + agents - Microsoft Software Engineer - CoworkBas Brekelmans
Software Engineer at Microsoft, started the Copilot Cowork project and previously CTO of Copilot Studio. The clearest voice on why Microsoft chose multi-model, MCP, and A2A architecture across the agent stack.
Microsoft Software Engineer, Cowork - ex-CTO Copilot Studio - Microsoft Principal PM - Copilot Studio (Power CAT)Henry Jammes
Principal Program Manager for Copilot Studio on the Power CAT team at Microsoft. Deep on agent-building mechanics, governance, and the release roadmap - a first-party source on what Copilot Studio can actually do.
Microsoft Principal PM - Copilot Studio (Power CAT) - Microsoft Partner Director - Foundry Agent ServiceJeff Hollan
Partner Director of Product at Microsoft, leading the Foundry agent platform: Agent Service, the Agent Framework, and the SDKs. The clearest source on how agents actually go from prototype to a hosted production runtime.
Microsoft Partner Director - Foundry Agent Service + Agent Framework - Microsoft Principal Group PM - DataverseJulie Koesmarno
Principal Group Product Manager at Microsoft, leading the Dataverse as Agent Data Platform team. Follow her to see where the governed data layer, MCP surface, and coding-agent plugins are heading before they ship.
Microsoft Principal Group PM - Dataverse as Agent Data Platform - Copilot Studio MVPLisa Crosbie
6x Microsoft MVP, Practical AI for Business with Copilot and Agents at Barhead. Calm, hands-on voice on production-grade agent design - the one to follow when you want patterns you can ship, not just demo.
6x Microsoft MVP - YouTuber + International Speaker - ~21K followers - Microsoft Principal Cloud Advocate - Power Platform Advocacy LeadApril Dunnam
Principal Cloud Advocate and Team Lead for the Power Platform Advocacy team at Microsoft, and the face of Agent Academy. If a Copilot Studio feature or governance update ships, April has usually broken it down within 48 hours.
Microsoft Principal Cloud Advocate - Power Platform + AI Demystifier
Coming up
- June 29, 2026 · EU Council formally adopts the Digital Omnibus on AI, the last step before publication in the Official Journal.
- July 1, 2026 · Copilot Cowork billing begins. The Frontier grace period ends - set Copilot Credit spend limits first.
- Coming weeks · The fine-tuned Cowork 1 model arrives, to run routine tasks at lower credit cost than Opus 4.8.
- End of June 2026 · Foundry Hosted Agents reach general availability. Still in preview this week; GA forecast for end of June into early July.
- August 2, 2026 · Core AI Act provisions take effect. The Omnibus postponed the high-risk dates, not these - GPAI and prohibited-practice rules still start on schedule.
This week's question
June 16 turned the meter on, and the Frontier grace period ends July 1. Two questions for Monday: have you set per-group Copilot Credit spend limits before the first metered invoice, and can you see credits-per-agent in the new Copilot Agent Kit analyzer yet? And the harder one: if a model behind one of your agents vanished tomorrow, the way Fable 5 still has not come back, how fast could you swap it?
Get Agentic Weekly in your inbox
Every Monday morning. Unsubscribe any time.