Home
Apr 23, 2026
View All

How Anthropic's PM Team Moves Fast and Microsoft Goes Full Agent Mode

·2 underrepresented voices

The Short Version#

Two complementary signals today: Anthropic's Head of Product for Claude Code reveals the internal practices that let their PM team move faster than nearly anyone in the industry — and Microsoft ships Agent Mode as the default in Word, Excel, and PowerPoint, marking the clearest sign yet that "AI assistant" is over and "AI agent" is the new baseline.

Lenny's Newsletter / Cat Wu — How Anthropic's Product Team Moves Faster Than Anyone Else#

Source: https://www.lennysnewsletter.com/p/how-anthropics-product-team-moves Credibility: High (direct interview with the Head of Product for Claude Code, published April 23, 2026)

What happened: Cat Wu, Head of Product for Claude Code, sat down with Lenny Rachitsky to walk through how Anthropic's product team actually operates. The episode covers how they build products before the model is ready, how they run their "launch room" process for Claude Code, and why speed has become the organizing principle of their PM culture. This isn't a puff piece about Anthropic — it's a working PM explaining specific practices that produce fast, high-quality output.

Key PM craft patterns:

  • Build before the model is ready: Anthropic PMs design and prototype against capability curves they expect to hit, not capabilities that exist today. This means product decisions are being made with expected future model behavior in mind — a fundamentally different discipline than building against a stable API
  • Launch room process: Claude Code uses a dedicated launch room practice where the PM and team maintain continuous presence through a launch window, triaging real-time feedback and making rapid call decisions. This is closer to incident management than traditional launch coordination
  • Speed as PM discipline: Wu frames speed not as a byproduct of a small team but as an explicit practice — PMs are expected to compress decision cycles, reduce async back-and-forth, and make calls with less information than they might otherwise want
  • AI is changing the PM role itself: The episode addresses directly how AI tooling is changing what PMs do day-to-day, with Wu noting that the skills that made PMs effective in slower cycles may not map cleanly to the current pace

Why it matters for PMs: If you're at a company that ships AI features but still runs a traditional quarterly planning and launch cadence, this episode is going to feel like a report from a different industry. The "build before the model is ready" pattern is particularly important — it implies that the best AI PMs are developing a new skill: reasoning about capability trajectories, not just current-state requirements. The launch room model is also worth stealing: it's a practice that treats AI launches as higher-variance events requiring tighter feedback loops than normal releases.

Critical questions:

  • How do you write a PRD for a product whose core capability doesn't exist yet? What are the failure modes when your capability assumptions turn out to be wrong?
  • The "speed over process" framing works at Anthropic because they have a small, senior PM team. Does it scale to larger orgs with more junior PMs or more stakeholder coordination overhead?
  • What does "AI is changing the PM role" actually mean in practice — is it changing what PMs do, or how fast they do it, or something more structural?
  • Wu is a woman leading one of the highest-profile PM roles in AI right now. What does her framing of the PM role reveal about what Anthropic actually values in product leadership?

Action you could take today: Listen to this episode and write down the three practices Cat Wu describes that you're not currently doing — then pick one and design a micro-experiment to test it with your team in the next sprint.

Microsoft — Agent Mode Is Now the Default in Word, Excel, and PowerPoint#

Source: https://www.linkedin.com/posts/satyanadella_copilots-agentic-capabilities-in-word-excel-activity-7452870230600548352-fMvs Credibility: High (Satya Nadella's direct announcement post, April 22–23, 2026)

What happened: Microsoft made Agent Mode generally available and set it as the default experience across Copilot in Word, Excel, and PowerPoint. This is not an incremental Copilot update — it's a positioning shift. Nadella's framing is explicit: "as models become more capable, we're bringing that power to where real work happens, right in the canvas." A single prompt can now "reshape the model, the bridge and the narrative at once." Separately, Microsoft also announced "Hosted agents in Foundry" — dedicated enterprise-grade sandboxes for agents with durable state, built-in identity, and governance support.

Key technical details:

  • Agent Mode is now the default, not an opt-in. Users opening Word, Excel, or PowerPoint encounter agent-first Copilot, not assistant-first
  • Agents operate "in the canvas" — meaning they reason over the document/spreadsheet context directly, not through a sidebar or chat pane
  • Hosted agents in Azure Foundry get their own dedicated compute sandbox with durable state — meaning agent state persists across sessions, a significant shift for enterprise workflow automation
  • Foundry agents have built-in identity and governance, making them auditable and policy-compliant without custom integration work
  • The framing: "Every agent will need its own computer" — Nadella is positioning Foundry as the infrastructure layer for enterprise agent deployment

Why it matters for PMs: Two things are happening simultaneously that PMs should pay attention to. First, the default shift: when a product changes its default, it's a statement about what the company believes users should experience — not what they optionally can. Making Agent Mode the default in Office means Microsoft has decided the average knowledge worker is now the agent user. That's a massive distribution event. Second, the Foundry "durable state" announcement solves a real enterprise problem: agents that forget everything between sessions aren't deployable in workflows that require continuity. Durable state + identity + governance is the enterprise readiness stack for agents.

Critical questions:

  • Agent Mode as default assumes users are ready for it. What happens to the large percentage of Office users who just want to write a Word doc? Is there user research supporting this default change, or is it a capability-first push?
  • Durable state is powerful but creates new failure modes — what happens when an agent's persisted state becomes stale or incorrect? How does Microsoft handle agent state rollback?
  • The "every agent needs its own computer" framing is a Foundry sales pitch. What's the actual cost structure here for enterprise customers, and does it change the build-vs-buy calculus for companies already using LangChain or other orchestration frameworks?
  • Microsoft's agent governance story is being told at the infrastructure level (Foundry). But governance for enterprise agents is also a product-level problem. Where does Foundry's responsibility end and the product team's begin?

Action you could take today: If you're building any workflow that touches Microsoft 365, read the Foundry hosted agents announcement and map out whether durable state changes the automation use cases you've scoped out or deprioritized. This might unblock something on your backlog.

OpenAI — ChatGPT Workspace Agents Ship for Teams#

Source: https://openai.com/index/introducing-workspace-agents-in-chatgpt Credibility: High (first-party OpenAI announcement, April 22, 2026)

What happened: OpenAI shipped Workspace Agents in ChatGPT — Codex-powered agents that automate complex workflows, run in the cloud, and connect tools across a team's environment. These are not single-session assistants; they're persistent agents that can be assigned work, run asynchronously, and report back. Separately, OpenAI also published a technical deep-dive on how they used WebSockets and connection-scoped caching in the Responses API to reduce overhead and improve latency for agentic loops — unusually transparent infrastructure writing.

Key capabilities:

  • Workspace Agents are Codex-powered and run in the cloud — meaning they execute code and interact with tools, not just generate text responses
  • Designed to "automate repeatable workflows" at the team level — positioned as team infrastructure, not individual productivity
  • "Connect tools and streamline team operations" — implying integration with external services, not just ChatGPT-native capabilities
  • The WebSockets addition reduces round-trip overhead in multi-step agentic loops, which is directly relevant for teams building agents on the Responses API

Why it matters for PMs: OpenAI is running the same play Microsoft just ran: moving from individual assistant to team-level agent infrastructure. The Codex foundation matters here — it means these agents can actually do things (write code, query APIs, manipulate data) rather than just answer questions. The WebSockets technical post is worth reading if your team is building on the Responses API; it shows OpenAI actively optimizing for multi-step agent loops, which signals where their engineering investment is going. The combination of team-scoped agents + infrastructure optimization is a clear bet that enterprise agentic workflows are the next growth vector.

Critical questions:

  • What does "connects tools securely" actually mean in practice — what's the permission model, and who controls what the agent can access?
  • How does Workspace Agents overlap with or replace the existing Assistants API? Is this a replacement, a layer on top, or a separate product entirely?
  • The Codex-powered framing positions these as coding-capable agents. But the "automate repeatable workflows" use case is much broader. Which users is OpenAI actually targeting first?

Action you could take today: If your team uses ChatGPT Team or Enterprise, test whether Workspace Agents is available in your org and run one of the example automations from the announcement post. Map it to a workflow your team actually does.

Quick Hits#

  • Lara Hogan: "Why can't they just...? Revisited" — a follow-up to her original post on why cross-functional frustration happens in teams, with updated thinking on org dynamics. Relevant for PMs navigating multi-team AI product work (2026-04-22): http://larahogan.github.io/blog/why-cant-they-just-revisited/

  • Karri Saarinen (Linear): Published AI impact data showing 80% increase in resolved work Q4 2025–Q1 2026, with Linear's own agent solving 57% of bugs reported in April. This is one of the first concrete productivity numbers from a product-led company using their own AI tooling in production (2026-04-17): https://www.linkedin.com/in/karrisaarinen/

  • OpenAI: Made ChatGPT for Clinicians free for verified U.S. physicians, nurse practitioners, and pharmacists — supporting clinical documentation, care, and research. Direct signal on OpenAI's vertical market strategy and willingness to use pricing as a distribution mechanism in regulated industries (2026-04-22): https://openai.com/index/making-chatgpt-better-for-clinicians

  • Simon Willison: Short post quoting Maggie Appleton — likely related to Appleton's ongoing writing on AI and human-computer interaction (see her "zero alignment" piece tracked earlier this week). Worth checking if you followed that thread (2026-04-23): https://simonwillison.net/2026/Apr/23/maggie-appleton/#atom-everything

  • Simon Willison: Writeup on Qwen3.6-27B, a 27B dense model claiming flagship-level coding performance. Relevant for teams evaluating open-weight models for coding tasks — smaller, cheaper, potentially deployable on-prem (2026-04-22): https://simonwillison.net/2026/Apr/22/qwen36-27b/#atom-everything

The Thread#

The default is becoming agentic. Three items this week tell the same story from different angles: Microsoft shipping Agent Mode as the default in Office, OpenAI launching Workspace Agents for teams, and Cat Wu explaining how Anthropic PMs already operate as if agents are the unit of work. The transition from "AI assistant as opt-in feature" to "agent as baseline experience" is no longer a roadmap item — it's shipping. The PM question shifts from "should we add an AI feature?" to "what's the right level of autonomy for the workflow we're designing?"

Sit With This#

Cat Wu describes Anthropic's practice of building products before the model is ready — designing against capability curves they expect to hit, not capabilities that exist today.

For your product: Is there a feature you've been waiting to spec until the AI capability is "good enough"? What would you build differently if you assumed the capability would be 2x better in six months — and what decisions would you make now to be ready for it?