Journal / Product / The Bottleneck is moving
The product operating system, now with agents
Every strong product company has a way of working it can describe. Amazon has Working Backwards. Spotify had squads and tribes. Basecamp wrote down Shape Up. Marty Cagan pulled the common thread together as the product operating model in Transformed: empowered teams that own problems and outcomes, not projects and dates.
What a product operating system has always done
Strip away the names and every product operating system runs the same loop:
- Collect signals. Customer calls, support tickets, sales notes, usage data, the market.
- Decide what matters. Group the signals, size the problem, compare it with the strategy.
- Define the bet. A problem brief, a PR/FAQ, a pitch or a PRD.
- Build and test. Prototype, validate with customers, ship.
- Take it to market. Release notes, help articles, sales enablement.
- Learn. Measure the outcome and feed it back into step 1.
Amazon's version starts at step 3 with a press release written before anything is built. Shape Up puts its weight on step 3 too, with fixed time and variable scope. Spotify focused on the team structure that runs the whole loop. The loop is the same; the emphasis differs.
The weak point was always the hand-offs. Signals sat in Zendesk and Gong. The PRD was out of date before engineering started. Release notes were written the night before launch. Each step needed a person to move the work along, and every hand-off lost context.
What changes with agents
Agents can now do the mechanical work at every step of that loop. They read every call transcript instead of a sample. They draft the brief from the evidence and keep it linked to the source. They build a working prototype on the real codebase. They write the release notes from the merged pull requests.
That does not make the people optional. It moves their work. Deciding which bet to take, judging whether the evidence is good enough and saying no stay human. The bottleneck moves from doing the work to directing it.
The companies packaging it
A new category of tools is turning this loop into a product. A few examples:
- Samepage Bets calls itself an operating system for product teams in the AI era. It pulls feedback from calls, tickets, CRM and Slack into one inbox and scores it. Agents then research, write specs, open pull requests in GitHub and draft release notes and help articles, with people reviewing each step on a shared board. It is the most complete version of the full loop I have seen.
- Devplan turns feedback from calls, tickets and Slack into ranked projects, PRDs and Jira or Linear tickets that engineers and coding agents can build from.
- Allstacks Product Studio focuses on spec quality. It grounds specs in the codebase, customer feedback and delivery history, and tests them with adversarial AI reviewers before work starts.
- BuildBetter covers the evidence side and passes that context to coding agents such as Claude Code, Codex and Cursor.
Most of these stop at the spec or the ticket. Samepage is one of the few that goes all the way to launch content.
Building it yourself in Claude
One of Samepage's own customer quotes says they first built something similar for themselves in Claude. That is a fair point of comparison, because most companies already have more of the pieces than they think.
Every product team has its own way of doing the separate parts of the loop: how it scores feedback, what goes into a problem brief, how it writes release notes, how it runs a usability test. Many teams have already turned some of these into Claude skills, short instruction files that Claude follows every time. What is usually missing is not the parts, but what connects them.
There are two ways to connect them:
- Chain the parts step by step. Connect each skill to the next, and use scheduled tasks for the steps that should run without anyone asking, such as a weekly digest of new themes from support and sales calls.
- Let Claude orchestrate. Claude already has access to the sources through connectors: calls, tickets, Slack, the codebase and your documentation. Instead of wiring every hand-off, you describe the whole loop once and let Claude decide which skill to run at each step.
At Swat.io we use the second option. Notion holds the documentation and the guide: how our product operating system works, which steps it has and which skill belongs to each step. Linear holds the work itself: the projects, issues and decisions. Claude reads the guide in Notion, gathers the evidence from the connected sources, runs the right skill and records the result in Linear, where the team reviews it and makes the call.
The skills themselves we add step by step. We started with prototyping on a copy of our front end with mock data, and usability tests on those prototypes that return results within two days. Earlier articles in this series describe both.
Buy or build
The dedicated tools win on team use and traceability. A board that engineering, product marketing and leadership all look at is hard to recreate with individual chats. Building it yourself wins on fit and ownership. Writing down how you work as skills forces you to make your operating system explicit, which most teams have never done.
Where to start
- Write down your loop. One page: which steps you run, who owns them and what each one produces.
- Pick the step that hurts most. For most teams this is turning scattered feedback into a clear problem.
- Automate that step in Claude first. Connect the sources, write one skill and run it for a month.
- Add the next step. Brief, prototype, release notes. One at a time.
- Evaluate a dedicated tool once the process works. When several teams need to run the same loop on a shared board, buying usually makes more sense than maintaining it yourself.
The product operating model was always the right idea. What held it back was the effort it took to run. That effort is now mostly gone, which leaves the part that was always the job: choosing the right bets.
If you run a product operating system with agents, bought or built, I would like to hear how you set it up.
