Shipping Got Cheap. Now Deciding What to Ship Is the Whole Job.

The old rule was simple. Shipping cost you a sprint, so you validated before you built. That rule is gone. When an agent can ship your idea within the hour, the bottleneck moves. It stops being "can we build it" and becomes "should we, and which one first."

A batch of new guides landed this month, all pointing at the same shift from different angles: how discovery works now, what AI PMs get grilled on in interviews, and how the coding tools your team relies on actually behave. Let me catch you up on what changed and what to do about it.

When building is the cheapest experiment you have

The math flipped. Paweł Huryn, who wrote his first discovery guide in 2022, puts it bluntly: back then, shipping unvalidated ideas was too wasteful. Now it is the opportunity cost. A team that shipped 2 bad ideas per sprint can ship 20 a week, plus the regressions users feel. Speed multiplies waste, so deciding what to ship became the job.

His rule of thumb is worth stealing. For cheap, reversible ideas, ship behind a feature flag and measure real usage. Building it can be the test. For risky or hard-to-reverse ideas, run a real experiment first, same as always. He points to one project with 1,082 unit tests that run on every push, where he shipped five releases in a single day. That only works because the safety net catches the mess.

So discovery is not dead. The output changed. It no longer hands off a validated backlog. It answers a sharper question: what is all this speed for?

The trio now includes teammates who don't sleep

The old lineup was a PM, a designer, and an engineer doing discovery together. Those role lines are blurring. Boris Cherny, who created Claude Code, described his own team as engineering, product, design, and data science melting into a new kind of role that is not tied to job function.

Agents are joining that room too. One synthesizes interviews, one builds the prototype, one wires up analytics. The catch is that an agent does bad work with a bad brief, same as a person. Hand it a task list with no context and you get noise. Give it an objective, the outcome you want, and the constraints, and you get something useful.

Keep one thing human on purpose: the customer conversations. Faster shipping only produces more noise if you do not understand the problem. Map where users actually struggle before you decide what to build next.

The interview room caught up fast

If you are hiring AI PMs, or you are one, the bar moved. Aakash Gupta went looking for content on AI PM technical rounds and found nothing, so he built the guide from real reports. His count: 12,397 AI product jobs posted in the US in the first half of 2026.

The questions are specific and technical. OpenAI asks you to define hallucinations in LLMs. Anthropic asks how you would define a model "redline." Nvidia asks you to design a RAG system on their infrastructure with latency, relevance, and cost tradeoffs. Perplexity asks you to explain how RAG works, full stop.

The pattern for your team: PM candidates now need real depth on how these systems work, not just a story about outcomes. If your interview loop for product roles still skips the technical round, you are screening for the wrong thing.

The context bill nobody reads

Here is the part that hits your engineering costs directly. Every tool you plug into an agent gets dumped into its context on every request. Sarath Krishna Prasad shows how a few busy MCP servers can burn tens of thousands of tokens on tool schemas before the conversation even starts. Google Calendar's event list alone exposes 20-plus properties when a simple question needs 5.

His fix is almost dull, and that is the point: give the agent a shell and good CLI tools. The model already knows git, gh, jq, and curl from its training. A CLI costs near zero tokens until the agent runs --help. When it fails, you see the exact error and can rerun the command yourself. MCP still earns its keep where there is no shell, or should not be one, like non-technical users or locked-down enterprise setups.

The deeper lesson from Software Workaholic's Claude Code writeup: more context is not better. The window grew from 200K to 1M tokens, and the instinct to celebrate was wrong. As it fills, the model loses the thread and output drops. They now clear and start fresh far more often than expected.

The deep cut

The habit that changes your Monday: stop treating the agent like a search box and treat it like a teammate you brief. That writeup lays out a five-step pipeline built on Matt Pocock's skills: grill the idea against real docs, write a spec, break it into tickets, implement one at a time, then run a dedicated code review. Thinking happens before building, and every stage produces an artifact a human reviews.

The money detail is model choice. Start with the everyday model, escalate to the heavy reasoner only when it fails. Defaulting to "always the biggest model, max effort" burns your limits fast for no gain. When a model gets lazy and refuses, that is often low effort, not too small a model. Match the tool to the task on purpose, and your bill and your quality both improve without new headcount.

Three questions for your team

1. Which of our next 10 roadmap items are cheap and reversible enough to ship behind a flag instead of running a full experiment first? Sort them now.
2. When we brief an agent, are we handing it a task list or an objective with context and constraints? Pull up your last three prompts and check.
3. Are we paying a token tax on MCP tools we barely use, when a plain CLI would cost near zero until called? Audit what is loaded into context on every request.