The Product Desk

The Throughput Numbers That Break Your Planning Model

Three independent data points from elite engineering organizations have converged into a single, unavoidable conclusion: coding agent throughput is no longer linearly tied to headcount. A 3-person OpenAI team built a million-line internal product in 5 months with zero hand-written code, averaging 3.5 PRs per engineer per day. Stripe's internal agents ('Minions') now produce 1,000+ merged PRs per week via unattended parallelization — a developer posts a task in Slack, the agent writes code, passes CI, and opens a PR with zero interaction. Solo developer Peter Steinberger made 6,600+ commits in a single month running 5-10 agents simultaneously.

If a competitor can stand up a 3-person team that outputs at the rate of a 15-person team, your feature parity timeline just compressed dramatically.

Mitchell Hashimoto (creator of Terraform) coined the term 'harness engineering' for this emerging discipline: building the constraints, documentation, and feedback loops that keep agents productive. The key insight is counterintuitive — you increase agent reliability by constraining the solution space, not expanding it. OpenAI enforced strict layered architecture with custom linters whose error messages double as remediation instructions. Stripe built Toolshed, a centralized MCP integration exposing 400+ internal tools to agents in sandboxed devboxes.

The Hard Reliability Ceiling You Must Design Around

Here's where the enthusiasm meets physics. Anthropic's Opus 4.6 benchmarks reveal a steep, quantifiable decay curve: 80% accuracy on 1-hour tasks, plummeting to 50% at 14.5 hours of autonomous work. Separately, Anthropic found agents marking features as complete without proper end-to-end testing — their agents couldn't even see browser-native alert modals with Puppeteer. OpenAI acknowledges that agent-generated code accumulates 'entropy' that looks different from human technical debt.

These two data sets — extraordinary throughput and steep reliability decay — aren't contradictory. They define the design envelope. The winning architecture is clear: decompose complex workflows into sub-1-hour atomic tasks, insert human checkpoints at each boundary, and design your UX around graceful degradation. Greg Brockman recommends every team designate an 'agents captain' responsible for making agents effective. Engineers cap out at 3-4 parallel agent sessions before becoming the bottleneck themselves.

The Greenfield Constraint Is a Strategic Lever

Every success story involves either greenfield projects or purpose-built harnesses. Retrofitting to legacy codebases is explicitly called out as unsolved. This creates a strategic opportunity: your next major feature or service can be architected from day one for agent-native development — strict interfaces, comprehensive AGENTS.md, MCP-exposed tooling, automated testing. The compounding nature of harness engineering means starting one quarter earlier creates a durable advantage.

OpenAI's Margin Collapse Is Your Cost Risk

OpenAI's leaked financials contain the most important data point for any PM building AI-powered products. Forget the headline revenue ($13.1B, slightly above projections). The story is in the margins: gross margin collapsed to 33%, a full 13 percentage points below their own 46% forecast. Model operating costs quadrupled in 2025. Forward projections show $111B in cumulative cash burn through 2030 — more than double their previous estimate. Their 2030 plan assumes a $28B training cost reduction to offset rising inference costs, a bet requiring massive technical breakthroughs.

If your product's AI roadmap has a single point of failure on OpenAI, you're carrying more risk than you realize.

Meanwhile, OpenAI is spreading itself thin: a $200-$300 smart speaker entering the Amazon Echo segment, a stalled Stargate infrastructure project, and an ongoing arms race with xAI (which merged with SpaceX and is heading for an IPO). Short seller Jim Chanos noted these five-year AI forecasts are 'just guesses' — and the guess keeps getting worse.

The Infrastructure Bottleneck Beneath the Model Layer

The binding constraint on AI product development is shifting from model capability to physical compute availability. New AI campuses are designed for 1 gigawatt+ of power — 10x traditional data centers. The talent to build at this scale is so scarce that data center executives command $10M+ packages, and lenders are inserting 'key man' clauses into billion-dollar project financing — meaning a single executive departure can trigger withdrawal of funding. When your cloud provider promises capacity in Q4 2026, that promise may be contingent on whether one specific person stays in their job.

The optimistic read: the sheer capital being deployed means compute costs are almost certainly coming down in 18-24 months. Features economically unviable at today's inference costs may pencil out soon. Smart PMs are doing design and prototyping work now on those features.

Enterprise Buyers Aren't Actually Switching — Yet

Here's the tension worth surfacing: SaaS stocks are down 20-30% YTD on AI disruption fears, but enterprise buyers aren't actually moving. Lead Edge Capital's Evan Skorpen says large companies 'have no interest in trying some vibe coded solution' and software is a tiny percentage of operating expenses. The switching cost math doesn't work. If you're an incumbent PM, your defensibility is stronger than your stock price suggests. If you're an AI-native challenger, stop leading with 'cheaper and AI-powered' and start leading with capabilities that don't exist yet.

Cisco's SVP Just Told Enterprise CIOs to Stop Buying Your Product

Cisco's SVP of AI Platform DJ Sampath delivered the clearest articulation yet of how enterprise procurement is being coached to evaluate AI products: 'Companies that are a thin shim on top of a model — their days are numbered.' He elaborated: 'Adding a generative API to an existing product isn't a strategy — it's a feature.' This was said at the Cisco AI Summit, attended by OpenAI, Nvidia, and Anthropic leadership. It's not a hot take; it's the framing being used to sell Cisco's full-stack AI platform to CIOs.

Simultaneously, OpenAI is building native platform connectors (e.g., HubSpot via GPT-5), signaling aggressive moves into the integration layer. HubSpot co-founder Dharmesh Shah confirms GPT-5 handles 80%+ of his AI interactions. If your AI differentiation is a well-designed prompt chain hitting Claude or GPT-4, enterprise buyers are being actively coached to see you as a feature, not a product — and your model provider is building the integrations that make you redundant.

The antidote: intelligence embedded into the product itself, trained on contextual enterprise data, improving continuously with each customer's usage.

The Security Gate You're Not Ready For

Sampath confirms attackers are already probing gaps in agentic AI systems — hijacking, impersonating, or manipulating agents to exfiltrate data at machine speed. MCP and agent-to-agent protocols 'have become the backbone of autonomous workflows but scaled faster than the security around them.' A Cambridge study found only 4 of 30 top AI agents have published formal safety evaluations. Browser agents — the most autonomous category — are missing 64% of safety disclosures.

Cisco's prescribed security stack — zero-trust identity for agents, behavioral monitoring, human-in-the-loop gates for privilege escalation — will likely become the enterprise procurement checklist within 12 months. The 70+ country Delhi Declaration signals regulatory frameworks are coming, while the U.S. rejection of global AI governance means the landscape will be fragmented. Publishing a formal safety evaluation now puts you in elite company and converts enterprise trust.

The 80/20 Design Pattern

Sampath predicts AI will resolve 80% of pattern-based, routine incidents autonomously within 12 months. The remaining 20% — multi-vendor, legacy-heavy edge cases — will take longer. This is a transferable design pattern: don't ship 100% autonomy. Design for the 80% that's routine, and invest your UX budget in making the 20% human-in-the-loop experience exceptional. Only 28% of organizations believe they're AI-ready, and the blocker is infrastructure debt, not model capability. Products that meet enterprises where they are — messy, fragmented, legacy-heavy — will capture the 72% that current solutions ignore.

LinkedIn's DPH Framework Gives You the Metrics You've Been Missing

LinkedIn open-sourced their Developer Productivity & Happiness (DPH) Framework — a complete system of metrics, processes, and feedback loops for understanding developer needs. For PMs, this matters because every roadmap conversation eventually becomes a capacity conversation, and capacity conversations without shared metrics devolve into opinion wars. The DPH Framework provides data-backed answers to 'why is this taking so long?' and 'where should we invest in tooling?'

This arrives at a moment when productivity measurement is more important than ever. With harness engineering enabling 5-10x throughput on greenfield work, the gap between AI-augmented and non-augmented engineering velocity is widening fast. Staff engineer Sean Goedecke reports that many engineers who don't get value from LLMs are 'holding it wrong' — not using models in the most effective ways. If half your team ships 30-50% faster with AI assistance and the other half isn't using it, your sprint planning is based on averaged velocity that doesn't reflect reality.

You're either over-committing because your fast engineers can't compensate enough, or under-committing because you're not accounting for AI-augmented speed.

The Cultural Signal Beneath the Data

Multiple sources surface a consistent finding: engineers who love solving algorithmic puzzles struggle to go agent-native, while those who love shipping products adapt quickly. This is a hiring and team composition signal. Combined with the recommendation for 'agents captain' roles and the observation that harness infrastructure doesn't happen by accident, the picture is clear: the transition to AI-augmented development requires deliberate organizational design, not just tool adoption.

The meta-signal from engineering leadership content is also instructive. The topics resonating most with your engineering counterparts right now are: AI coding tool adoption, developer productivity measurement, async work practices (PostHog operates fully async across 11 countries), and the death of Scrum at scale in Big Tech. If you're trying to build stronger alignment with engineering leadership, these are the wavelengths they're on.