Home
May 22, 2026
View All

AI Creates More Work, Not Less — and Agent Infrastructure Gets Serious

·1 underrepresented voice

The Short Version#

Dan Shipper's report from inside Every is the most honest thing written about AI and headcount this year, LangChain shipped a deep series on agent infrastructure that every PM building agentic products should read, and Amazon Nova Act becoming HIPAA eligible is a quiet but significant unlock for anyone building AI agents in healthcare.

Dan Shipper / Every — AI Progress Creates More Work for Humans, Not Less#

Source: https://x.com/danshipper/status/2057514494960513272 and https://every.to/@danshipper Credibility: High (first-person operational data from a company that has been building with AI since GPT-3)

What happened: Dan Shipper published a report on what's actually happening to headcount when you go all-in on AI. Every has automated "every single thing we can" with AI agents. They've also grown from 4 to 30 human employees since GPT-3. His core argument: AI makes expert competence cheap, which drives up demand for expertise, which creates more work, not less. The productivity gains get absorbed by the expansion of what's possible.

Key patterns:

  • Automation doesn't reduce labor demand when the work itself is elastic. Making expert output cheaper expands the market for it.
  • The structural reason isn't that AI is inefficient. It's that lowering the cost of quality output raises the ceiling on how much quality output a team wants to ship.
  • This is Jevons Paradox applied to knowledge work — the efficiency gain gets consumed by increased consumption of the newly-cheap resource.
  • Every went from a tiny media operation to a 30-person company because AI made it viable to do things they couldn't afford to attempt before.

Why it matters for PMs: This is the answer to the question product leaders keep asking: "Should I worry that AI will shrink my team?" The more relevant question is what expands when your team gets more capable. Shipper's data suggests the answer is scope, not headcount reduction. For PMs building AI features: if your AI feature makes something cheaper, don't model it as cost savings. Model it as demand generation. Users will do more of the thing that got cheaper.

Critical questions:

  • Does this pattern hold only for creative/editorial work, or does it generalize to ops, engineering, and support contexts?
  • What's the lag between AI adoption and headcount expansion? Every's 4-to-30 happened over multiple years, not immediately.
  • Is this specific to companies that are building with AI vs. companies in industries disrupted by AI? The causal mechanism might not transfer.
  • At what point does the expanded demand plateau? Every is still growing, but there's presumably a ceiling.

Action you could take today: Take one AI feature you've shipped or are planning. Map out what users would do more of if the cost of that activity dropped by 80%. That demand forecast is probably more accurate than assuming the AI feature reduces support tickets or saves time.

LangChain — Agent Infrastructure Series: Streams, Sandboxes, and Harnesses#

Source: https://www.langchain.com/blog/token-streams-to-agent-streams, https://www.langchain.com/blog/the-anatomy-of-an-agent-harness, https://www.langchain.com/blog/how-auth-proxy-secures-network-access-for-langsmith-agent-sandboxes Credibility: High (first-party technical documentation from the team building the production infrastructure)

What happened: LangChain published a tight cluster of posts this week that together describe what production agent infrastructure actually looks like. Three posts cover: how streaming moves from token-level to typed agent events; what an "agent harness" is (filesystems, sandboxes, memory as the three components that turn a model into an autonomous work engine); and how Auth Proxy keeps secrets out of sandbox runtimes while controlling what agents can reach on the network. This is less a blog series and more a technical spec for teams building serious agent systems.

Key technical details:

  • Agent streaming now emits typed events, not just token strings. This means frontend subscriptions can be scoped to specific subagents or event types, and multimodal outputs (not just text) are first-class.
  • An agent harness has three required components: a filesystem (where the agent reads/writes work), a sandbox (isolated execution), and memory (what persists across runs). If any of these are missing, you don't have an autonomous agent — you have a chatbot with tool calls.
  • Auth Proxy sits between agent sandboxes and the outside world. Secrets never enter the sandbox runtime. All network egress is constrained. This is the infrastructure-level answer to "how do you let an agent do real work without it leaking credentials or making unauthorized calls."
  • Deep Agents v0.6 is the release that ships these streaming primitives into production.

Why it matters for PMs: If you're speccing or reviewing agent features, this gives you a vocabulary for what's actually required. "We'll build an agent" is not a spec. "We need a harness with a scoped sandbox, typed event streaming, and credential isolation via Auth Proxy" is a spec. The gap between those two descriptions is where most agent projects get stuck. LangChain is essentially publishing what the production checklist looks like.

Critical questions:

  • How much of this is LangGraph-specific vs. portable to other agent frameworks? Auth Proxy and LangSmith Sandboxes are LangChain products — the pattern may not transfer easily to teams building on other stacks.
  • The typed event streaming model requires frontend changes. What's the migration path for teams already running token-stream-based UIs?
  • "Memory" is glossed as "what persists across runs" — but the hard problem is what to persist and when to discard. Does the harness model address this, or is it left to the application layer?
  • What does the failure mode look like when Auth Proxy misclassifies a legitimate egress call as unauthorized?

Action you could take today: If your team is building or evaluating any agentic feature, use the three-component harness model (filesystem, sandbox, memory) as a checklist in your next design review. If any of the three is undefined or "TBD," that's your blocker to name explicitly before you spec the feature.

Simon Willison — FTC Settles "Active Listening" AI Marketing Case for $1M#

Source: https://simonwillison.net/2026/May/22/ftc-active-listening/#atom-everything Credibility: High (Simon is linking to the FTC enforcement action directly; this is a regulatory outcome, not a rumor)

What happened: The FTC required Cox Media Group and two other firms to pay nearly $1 million to settle charges that they deceived customers about an "active listening" AI-powered marketing service. The service allegedly claimed to use microphone data from users' devices to target ads. The FTC found this deceptive. Simon flagged it as a notable signal about where AI marketing claims run into regulatory walls.

Key patterns:

  • The enforcement is about deceptive claims around AI capabilities, not necessarily the underlying technology itself.
  • "Active listening" as a product positioning concept is now officially in the FTC's crosshairs. Any product that implies passive audio monitoring for advertising purposes is facing regulatory risk regardless of whether it actually works that way.
  • The $1M settlement is modest, but the FTC action creates precedent and signals the kind of AI feature claims that will draw scrutiny.

Why it matters for PMs: Two levels here. First, if you work on any product that uses audio, ambient sensing, or behavioral inference for targeting, this is a direct regulatory signal. Second, and more broadly: the FTC is now actively evaluating whether AI capability claims in marketing are accurate. "Powered by AI" in product copy is fine. Claims about what the AI can do — especially anything that implies user data collection or inference without explicit consent — is not.

Critical questions:

  • How does this interact with on-device AI (like Apple Intelligence) where inference happens locally without data leaving the device? Is the FTC's concern about data collection, or about the capability claim itself?
  • What's the disclosure standard the FTC is implicitly establishing? Is "we use your microphone data" in the terms of service sufficient, or is conspicuous disclosure required?
  • Does this affect ad tech companies using behavioral inference that doesn't involve audio? The settlement is narrow, but the FTC's theory could expand.

Action you could take today: Review your product's AI marketing copy for any capability claims that could be read as implying passive data collection. If you're at a company with an ad product or behavioral targeting feature, flag this settlement to your legal and marketing teams before your next campaign brief.

Amazon Nova Act — Now HIPAA Eligible#

Source: https://aws.amazon.com/blogs/machine-learning/amazon-nova-act-is-now-hipaa-eligible/ Credibility: High (first-party AWS announcement)

What happened: Amazon Nova Act, AWS's agent-focused model, is now HIPAA eligible. This means healthcare organizations can use Nova Act to build agentic workflows that handle Protected Health Information (PHI) within a covered entity relationship. The post walks through what HIPAA eligibility means for agentic AI specifically — the compliance boundary, how BAAs (Business Associate Agreements) apply, and how to get started.

Key details:

  • Nova Act is specifically designed for agentic use cases — it's optimized for multi-step task completion, not just single-turn responses.
  • HIPAA eligibility means AWS will sign a BAA for Nova Act usage, which is the contractual requirement for healthcare companies to use any third-party service with PHI.
  • This unlocks agentic AI for clinical workflows, prior authorization, care coordination, and administrative automation in healthcare — all contexts where multi-step autonomous action has obvious value but compliance has been the barrier.
  • The AdventHealth case study (also published this week, separate item) shows ChatGPT being used to reduce administrative burden in healthcare, so there's market validation for the category.

Why it matters for PMs: Healthcare AI has been stuck at "we can show you a demo but we can't touch real patient data" for a long time. HIPAA eligibility for an agent-capable model changes that calculus. If you're building in digital health, health tech, or any enterprise product that touches healthcare workflows, Nova Act's eligibility means the compliance conversation just got shorter. This is a build-vs-buy signal: you can now buy HIPAA-compliant agentic AI rather than building a compliance wrapper around a non-eligible model.

Critical questions:

  • HIPAA eligibility is necessary but not sufficient. Does Nova Act's performance on healthcare-specific tasks (clinical terminology, nuanced care coordination logic) hold up? The compliance box is checked; the quality bar still needs evaluation.
  • How does this compare to Anthropic's HIPAA-eligible Claude models for agentic healthcare use cases? Both are now available — the choice comes down to task performance and tooling integration.
  • What audit trail does Nova Act produce for agentic actions in a HIPAA context? Clinical compliance requires documentation of what an agent did and why.

Action you could take today: If you're a PM on a healthcare product that has been waiting on compliance coverage before piloting AI agents, this is your "start the conversation with your compliance team" moment. Forward the AWS post with the question: "Does this unblock our agent pilot?"

Quick Hits#

  • Andrej Karpathy joins Anthropic: Karpathy announced he's joining Anthropic for R&D, calling the next few years at the frontier of LLMs "especially formative." One of the most credible AI educators and researchers moving to Anthropic is a meaningful signal about where frontier research is happening. (2026-05-19): https://x.com/karpathy/status/[May-19-2026]

  • Character.AI — Smarter Memory: Character.AI shipped "Smarter Memory for Smarter Chats" — persistent memory updates in the consumer conversational AI product. Memory as a retention driver is a pattern worth tracking in consumer AI. (2026-05-21): https://blog.character.ai/memory/

  • Simon Willison — Datasette Agent: Simon shipped Datasette Agent, an agent that can query and visualize data in Datasette, plus companion plugins for sprites and charts. A working single-developer example of agent + data tool integration. (2026-05-21): https://simonwillison.net/2026/May/21/datasette-agent/#atom-everything

  • Teresa Torres — My Team of Agents: Teresa Torres wrote about running a personal team of Claude agents that complete tasks while she's away from the computer — a PM-practitioner's perspective on async agentic workflows. (2026-05-20): https://www.producttalk.org/my-team-of-agents/

  • Figma — Bulk Edit in Figma Buzz: Figma shipped bulk edit and resize for campaign assets in Figma Buzz — upload a spreadsheet, multi-select cells, manage content at scale. Small feature, but points toward Figma Buzz becoming a real creative production tool. (2026-05-22): https://help.figma.com/hc/en-us/articles/31271824185623-Bulk-create-assets-in-Figma-Buzz

The Thread#

The agent infrastructure story is getting concrete. LangChain's harness framework, Nova Act's HIPAA eligibility, Cursor's Jira integration, and Teresa Torres's personal agent setup all point to the same thing: agentic AI is moving from "interesting demo" to "things with real operational requirements." Compliance, credential isolation, typed event streams, task handoff from project management tools — these are not research problems anymore. They're product specs.

Sit With This#

Dan Shipper's data shows Every grew from 4 to 30 employees as AI automation increased — not despite it, because of it. The productivity gains unlocked scope expansion faster than they reduced headcount.

For your product: Pick an AI feature you've framed internally as a cost savings or efficiency play. What would it look like to reframe it as a demand generator instead? What would users do more of if the cost of that activity dropped significantly — and is your product positioned to capture that expanded demand?