Home
May 2, 2026
View All

Microsoft Agent 365 Goes GA and Open-Source Models Are the New Default

·1 underrepresented voice

The Short Version#

Two signals today point in the same direction: AI tooling is maturing from "cool capability" to "enterprise infrastructure." Microsoft flipped the switch on Agent 365 — a full control plane for managing agents at scale — and Harrison Chase is openly saying the economics of closed models are broken, which accelerates the path toward OSS-first agent architectures.

Microsoft — Agent 365 Generally Available#

Source: https://www.microsoft.com/en-us/security/blog/2026/05/01/microsoft-agent-365-now-generally-available-expands-capabilities-and-integrations/ Credibility: High (first-party announcement from Microsoft Security blog)

What happened: Microsoft Agent 365 — their control plane for enterprise AI agents — shipped to general availability on May 1. This is the management layer that sits above individual agents and their interactions: identity, security, governance, compliance, and auditability across the enterprise. In Satya Nadella's framing, they're "extending the systems customers already use for identity, security, governance and management to every AI agent."

Key capabilities:

  • Centralized identity and access management for agents — every agent gets a managed identity, not just a login credential
  • Security and governance policies that apply across agents, not per-agent configuration
  • Audit and compliance trails for agent interactions (critical for regulated industries)
  • Integration with existing Microsoft security stack (Entra, Purview, Defender)
  • Works across agents running inside and outside the Microsoft ecosystem

Why it matters for PMs: This is the moment when "we have AI agents" becomes "we can actually run AI agents in production." Enterprise PMs have been blocked on agentic deployments by a simple question: "Who is responsible when an agent does something it shouldn't?" Agent 365 is the answer — it creates a governance layer that security and compliance teams can actually reason about. If you're building anything that enterprise customers will deploy at scale, this is now the integration surface you have to think about. Expect procurement conversations to start including "does this work with Agent 365?" the same way they included "does this work with Azure AD?"

Critical questions:

  • How does Agent 365 handle agents from non-Microsoft vendors — is this genuinely open or a soft lock-in play?
  • What's the observability model? Can you actually debug an agent's decision chain through the governance layer?
  • Does this apply to agents running in browsers or devices (client-side), or only cloud-hosted agents?
  • What's the pricing model — per-agent, per-interaction, or bundled into existing M365 tiers?

Action you could take today: If you're building an agent product targeting enterprise, pull up your current permission and audit model and map it against what Agent 365 expects. Find the gaps before a customer security review does.

Harrison Chase (LangChain) — OSS Models Are the 2026 Bet#

Source: https://x.com/hwchase17/status/2049552801890771220 Credibility: Medium (single tweet, but from the creator of LangChain — directly relevant to where agent infrastructure is heading)

What happened: LangChain's founder and CEO posted bluntly on April 29: "big theme of 2026 - cost of closed models is too high! really excited to make deepagents work exceptionally well with OSS models." This is a signal shift — LangChain built its reputation on being model-agnostic, but Chase is explicitly naming OSS as the direction for their deep agent product.

Key patterns:

  • Economics, not capability, is now the blocker for production agent deployments at scale
  • "deepagents" is LangChain's framing for multi-step, long-running agent workflows — where token costs compound fastest
  • This aligns with Meta's Llama 4 push and Mistral's continued open-weight releases — the OSS model ecosystem is now good enough to run production workloads
  • Closed model providers are facing margin pressure from their largest customers (developers building on top of them)

Why it matters for PMs: If you're doing the math on your AI product's unit economics, the token cost problem gets worse the more agentic your product becomes. A single "agent run" that involves research, planning, execution, and verification can easily be 50–200x the token cost of a simple completion. Chase naming this as the defining theme of 2026 tells you that the most sophisticated builders are hitting this wall right now. The practical implication: if you're designing agent workflows, your architecture decisions today are your cost structure in six months.

Critical questions:

  • What quality tradeoffs come with OSS models in agent workflows — where do they fail that closed models don't?
  • Is this actually an economics problem or a control problem (data privacy, compliance)?
  • How does "deepagents on OSS" affect the evaluation and observability burden?
  • If Anthropic and OpenAI respond with lower agent-tier pricing, does this argument collapse?

Action you could take today: Pull your last month's LLM API spend, break it out by feature/workflow type, and identify which features are driving the majority of token costs. That's where the OSS conversation starts.

Notion — Custom Agents in Private Slack Channels#

Source: https://www.notion.so/releases/2026-05-01 Credibility: High (official Notion releases changelog)

What happened: Notion shipped the ability to add Custom Agents to private Slack channels as of May 1. This extends Notion's agent capabilities — previously limited to Notion's own surfaces — into the communication workflows where decisions and follow-ups actually happen. Custom Agents can now be where the work is being discussed, not just where it's being documented.

Key capabilities:

  • Custom Agents (built on top of Notion's AI and knowledge base) can now be invited into private Slack channels
  • This means Notion's agent can respond to questions, pull context, and take actions inside Slack conversations
  • Private channel support is notable — this isn't just public/general channels, it's the sensitive project and team spaces

Why it matters for PMs: Notion has been quietly building toward being a knowledge layer that extends into other tools, not just a destination you go to. Adding agents to private Slack channels is the clearest version of that yet — now Notion's "brain" is where your team already works. The adoption pattern to watch: teams that previously had to switch context to query Notion can now ask in-channel. This lowers the activation energy for AI-assisted information retrieval significantly. Worth noting this is also a moat-building move — the more Notion's agent becomes part of daily Slack workflows, the harder it is to replace.

Critical questions:

  • What actions can the agent actually take from Slack — read-only, or can it write back to Notion?
  • How is access scoped — can the agent see all of Notion, or only pages the Slack workspace has been authorized for?
  • How does this interact with Notion's existing enterprise data residency and compliance posture?

Action you could take today: If your team uses both Notion and Slack, add the Notion Custom Agent to one active project channel this week and measure how often it gets used vs. direct Notion queries over the next two weeks.

Quick Hits#

  • Stripe: "Everything we announced at Sessions 2026" — Will Gaybrick's recap of Stripe's annual conference, including making Stripe "more programmable" and agentic commerce updates. Worth reading if you're building anything in the payments space. (April 29, 2026): https://stripe.com/blog/product

  • Ben Tossell: "Building gets easier" — Ben's Bites post on Codex syncing memories and files to a cloud agent drive; hands-on experience with the emerging "cloud brain" pattern where agents have persistent, portable context. (April 29, 2026): https://www.bensbites.com/p/building-gets-easier

  • Ammaar Reshi: Built an on-device vibe coding app powered by Gemma 4 running locally on Mac with MLX — demonstrated building the Chrome Dino game entirely offline, then open-sourced it. Relevant signal for on-device AI and the growing case for local model deployment. (April 28, 2026): https://x.com/ammaar/status/2049169134429073471

  • Simon Willison: Notes on Codex CLI 0.128.0 adding a /goal command — a small but interesting UX pattern where the agent is given an explicit success criterion rather than just a task. Worth watching as a pattern for how to scope agentic work. (April 30, 2026): https://simonwillison.net/2026/Apr/30/codex-goals/#atom-everything

  • Teresa Torres (Product Talk): Interview with AITropos — an AI employee for hospitality that takes orders where customers already are (WhatsApp, SMS, etc.). Strong example of meeting users in existing channels rather than forcing new app adoption, directly relevant to the Notion/Slack integration pattern above. (April 30, 2026): https://www.producttalk.org/building-ai-employees-for-hospitality-how-aitropos-takes-orders-where-customers-already-are/

The Thread#

Enterprise AI is entering its "infrastructure phase." Three signals this week point the same way: Microsoft ships Agent 365 (governance and identity for agents), Harrison Chase says agent economics require OSS models (cost at scale), and Notion pushes agents into private Slack channels (meeting users where they already work). The pattern isn't "AI does cool new things." It's "AI is now something your IT, finance, and security teams have opinions about" — which is exactly what happens when a technology crosses from experiment to infrastructure.

Sit With This#

Microsoft Agent 365 shipped a governance and identity layer for enterprise agents — centralized audit trails, managed identities, compliance controls across every agent interaction.

For your product: If your product includes AI agents that enterprise customers deploy, what happens when a customer's security team asks "who authorized that action, and can you show me the log?" Can you answer that today? If not, is that a product gap you own — or are you betting on the platform (Microsoft, Okta, etc.) to own it for you?