The Product Desk

The Workflow That Shipped This Week

A PM at Notion opens a task doc, writes four sentences, attaches a screenshot, and @mentions Codex through an internal system called Boxy. Twenty minutes later she has a PR, UI verification screenshots, and a preview URL. This is not a sandbox. It is how Notion ships. In a different building, Jarred Sumner pointed AI agents at Bun and had them rewrite 960,000 lines from Zig to Rust in 6 days, passing the Linux test suites. The common constraint in both stories: the human owned architectural judgment, the agent owned execution speed. Separate the thing being pitched ("agents ship features") from the thing being done ("a human with a clear spec delegates bounded work").

What the Spec Must Contain

The four sentences that actually work name the user problem, the success condition, the constraints the agent should not violate, and the edge cases that matter. Notion keeps Markdown spec files in the repo written in plain English with code pointers and verification steps. Those specs are what the agent reads. They are the source of truth now, not the code. Vague requirements used to produce a conversation with an engineer. They now produce a shipped feature that is wrong in 20 minutes.

The PM skill of thinking clearly about what a feature should do became directly executable. Spec quality now determines shipping velocity in a measurable way.

The Infrastructure Dependency Most Teams Miss

Notion is cutting CI time to 25% of current because agents idle during CI. The arithmetic is blunt. A 60-minute CI run means one iteration per hour. A 3-minute run means 20x more iterations. CI investment stops being a developer-happiness line item and becomes a product-velocity line item. That is the framing PMs should bring to the prioritization meeting.

The Maintenance Question Nobody's Answered Yet

Here is where teams tell themselves a comforting story. Notion reports specs producing shippable code at high rates. DevOps research shows AI-assisted code carries higher flaw density than legacy code, and the per-commit maintenance cost of merged AI work is a metric almost nobody tracks. The honest diagnostic is time-to-second-touch: how long between a merge and the next change to the same file for the same reason. If that number trends down while AI commit volume trends up, the velocity chart is debt wearing a speed costume.

Where This Works vs. Where It Breaks

The 2x2 to take into Monday's planning. One axis: is the feature bounded (one screen, known data model, clear success state) or unbounded (billing, auth, anything litigable). Other axis: does the team have review infrastructure (tests, staging, rollback) to catch confident-but-wrong output. Bounded plus review infrastructure is the only cell where four-sentence specs earn their keep. Every other cell ships bugs to production at machine speed.

The Bifurcation Thesis

Ben Thompson's latest Stratechery piece identifies a structural split in the inference market that most product teams haven't priced for. 'Answer inference' serves humans staring at loading spinners — fast, expensive, premium GPU silicon. 'Agentic inference' serves autonomous job queues where nobody's watching — slow, cheap, commodity hardware. Thompson argues agentic inference "will be the largest compute market by far because it won't be limited by humans or time."

A chatbot stays on premium silicon because latency is the product. A background enrichment pipeline does not, and its per-token cost is going to drop as commodity hardware gets qualified.

The Infrastructure Signals Validating the Split

Three deals this month confirm the thesis. Anthropic leased SpaceX's entire Colossus 1 — 220,000+ GPUs, 300 megawatts — because demand outpaces provisioned compute (users are hitting Claude rate limits). The $1.8B Akamai deal over 7 years bets on distributed edge inference for latency-sensitive workloads. Meanwhile, Nvidia is shipping standalone memory racks, CPU racks, and RTX Pro chips (less powerful, cheaper) for edge data centers. The hardware vendors see the split and are building for both tiers.

What Teams Get Wrong

The dangerous configuration: token-scaling value sold on a per-seat contract. Agent workloads multiply tokens per task by a large factor. Cost per completed task is falling more slowly than cost per token. A PM whose feature spawns 40 model calls per user request is watching margin erode even as model prices drop — because usage grows into the margin faster than prices fall. Sources agree: teams that assume the cost curve alone will fix unit economics are going to discover heavy users found the loop first.

The Architecture Decision

For every AI feature, answer one question: does a human wait for this token, or does a job queue wait? If a human waits, keep it on the premium path and price for latency. If a queue waits, architect it for the commodity tier now — don't wait for the pricing to split at the API level. Anthropic's capacity expansion and Gemini 3.1 Flash-Lite at sub-second / 1.8s p95 give you the tiers to route between today.

The Escalation Curve

A pentester ran the same agent benchmark twice, a year apart. Success rate went from 6% to 81% in one year. In the same window, Google confirmed the first AI-discovered zero-day used by cybercriminals, a 2FA bypass exploiting what the writeup called a "faulty trust assumption." A Claude-powered Cursor agent deleted an entire company database plus all backups in 9 seconds, because Railway's backup lifecycle was coupled to resource deletion and nobody had separated the two. AI systems now show significantly higher severe-flaw density than legacy applications in penetration testing. These are measured outcomes in production, not slides.

ServiceNow Just Named the Category

ServiceNow and NVIDIA shipped Project Arc, an autonomous desktop agent governed by AI Control Tower, with open-source benchmarks for measuring enterprise AI agent performance. The thing being pitched is an agent. The thing being done is defining what "good" looks like before procurement does. Whoever writes the eval writes the RFP. Separately, SAP locked its APIs against third-party agents and pushed developers to SFTP and screen automation workarounds. The incumbents are drawing the governance layer this quarter.

Enterprise buyers are not asking for agent orchestration. They are asking who approved the agent to touch the ticketing system, what it did last Tuesday at 2 a.m., and whether the audit log will survive a compliance review.

The MCP Monitoring Gap

A security lead opens the SIEM and searches for agent activity. She finds user activity. MCP connections from AI agents to enterprise data are completely unmonitored in most deployments. An agent operating inside an authenticated session is, from the logs, indistinguishable from the user whose token it borrowed. Audit trails built for humans do not survive contact with agents making 50 actions per minute across parallel tabs. Enterprise CTEM programs cannot see those calls. This is a shadow AI problem the existing tooling does not address.

Three Concrete Requirements for Every Agent Feature

1. Granular permission scopes per agent capability — narrow enough that one broken capability does not take down the workflow next to it
2. Runtime authorization that can approve or deny in real time based on context, not a config set at deploy time and forgotten
3. Complete audit trails that distinguish agent actions from human actions, with a named remediation owner per capability

These are already appearing on enterprise procurement questionnaires. A PM shipping agent features without them will discover the requirements one of two ways: during a security review that blocks the deal, or during an incident that writes the policy. The Monday decision is which three capabilities get scoped, authorized, and audited this sprint, and which ones get pulled from the demo until they are.

The Research Finding

A developer writes with an AI assistant for ten minutes, then the extension fails. She stares at the editor. Carnegie Mellon, MIT, Oxford, and UCLA found that 10 minutes of AI assistance creates ~20% performance degradation when the tool is removed, and users are nearly 2x more likely to skip tasks entirely rather than attempt them unaided. The effect holds across experience levels, so this is not a junior-user story. Well-designed autocomplete drops the cost of the next keystroke to zero. Users behave exactly as the UX asked them to.

The Retention Trap It Creates

Teams read dependency as a moat. The dashboard cooperates. Engagement rises, time-to-output falls. What falls with it is the user's ability to judge whether the output is any good, and that number goes down with it. That is the variable that decides next year's renewal. A user who cannot function through an outage is one bad Tuesday from cancellation. The competitor who ships something that makes that same user feel sharper wins quietly.

The Psychological Headwind Most Teams Aren't Measuring

Over 60% of US workers cite AI as a major stress source. Office crying is up 12 points YoY. Gen-Z is building offline cyberdecks to escape AI homogeneity. 83% of executives say psychological safety boosts AI success, yet Meta employees are described as "miserable" under AI mandates. Executive enthusiasm and user resistance are pulling apart. Adoption numbers that look healthy are measuring anxiety-driven compliance.

Progressive disclosure, guardrails, and skill-preservation are the usual answers. The better question is which of those three the specific workflow actually needs.

The Design Split

Most shipped features sit in the "produces answer + user slower" cell. The cell that compounds is "scaffolds + user faster." Build assistance that compounds capability. Think guided practice with receding scaffolding, not always-on autopilot. Make the user do the last 10% of the work and refuse to do it for them.

The Counter-Signal: Entertainment Wins Retention

Janitor AI runs at 2.5M daily active users for fantasy roleplay and is the 10th most popular consumer AI app. Productivity AI gets opened when a task arrives. Companionship AI gets opened every day. If the product's job is retention, the dependency research applies. If the job is entertainment, it does not, and 2.5M DAU is the benchmark.