Generative · loading…

What's the deal with Software Factories?

A software factory is a mix of agents, skills, harnesses, LLMs, rules and context that produces and maintains software from an idea, feature description, or reported problem. These are currently active and operational underneath services that you and I use every day, including Uber, Linear, and Ramp.

In an Uber demonstration, a request in Slack set agents researching stadium pickups, developing designs, building a working prototype (go watch! It's only 20 minutes, and if you're not already up on this stuff, it's important to see the state of it).

Linear CEO Karri Saarinen argues that ambition explains much of the difference between companies adopting these methods. Designers, product managers and other non-engineers can investigate problems, build prototypes and submit fixes. He also recommends automations to keep references to product areas and code owners current, helping work reach the appropriate humans.

At Ramp, Inspect gives an agent an isolated development workspace and lets colleagues join a shared session to add context, flag constraints and steer work before review. A separate maintenance workflow begins when monitoring detects a problem. An agent investigates, reproduces the problem in an isolated workspace and proposes a fix; an engineer must review before the code is merged into the product’s codebase. Authority varies with the task, but generally this kind of work can proceed without waiting for approval.

Inside the software factory A synthesis of agent-assisted software development at Uber, Ramp and Linear. Direction and delegated scope, acceptance of changes, and authorization of customer exposure govern a five-stage lifecycle: plan and design, build, verify, release, operate and learn. Failed checks loop from verification back to building. Results and customer feedback create new work. Shared foundations supply models, tools, access, context, skills, isolated workspaces, telemetry and cost measurement. Specific review and release rules vary. Inside the software factory Agents working across the software development lifecycle DIRECTION & AUTHORITY · PEOPLE & POLICY Set direction & delegated scope Accept the change Authorize customer/ business exposure Authority may be delegated in advance Test, repair & retest 1 PLAN & DESIGN Find context Draft requirements Explore prototypes 2 BUILD Write code in an isolated workspace Use shared tools 3 VERIFY Run tests & checks Review code & evidence Propose the change 4 RELEASE Merge accepted code Stage the rollout Control exposure 5 OPERATE & LEARN Watch production Investigate alerts Propose repairs Results, incidents & customer feedback SHARED FOUNDATIONS THROUGH A ~SINGLE GATEWAY Models Tools & access Code & organizational context Reusable skills Isolated workspaces Telemetry & costs UBER · RAMP · LINEAR / NOVEMBER 2026 CPJ.FYI / RADAR

Uber’s (hypothetical, sanitized, simplified) World Cup walkthrough moves through research into previous events and stadiums, North American requirements, two design variants, an experiment plan and a prototype built using existing software. Multiple agents assist the team along the way. Shared access controls, organizational context, reusable instructions, development workspaces and tests are automatically accessed (and in many cases, token-optimized) where needed. That foundation lets each new workflow use capabilities the company has already assembled.

Taking this demo literally assumes a lot of business logic is delegated to the factory floor (potentially via self-management?).

Who owns the World Cup initiative? Who can say yes to a VIP offer, commit the budget for driver incentives or change each locality’s plans?

In some companies, launching this idea directly from the factory would cause a political crisis: people have permission to Build and Ship; permission to Bet sits across several functions. Steering is somewhere else entirely.

Level The company is saying… What you’re allowed to do

Build

“Show us something.”

Make prototypes, investigate problems, demonstrate possibilities.

Ship

“Make it better.”

Put improvements into production within the product’s existing purpose and commitments.

Bet

“Try something new.”

Run live experiments, spend an agreed budget and test new offers with customers.

Steer

“Choose where we go.”

Set priorities, allocate resources and decide which opportunities the business pursues.

Linear describes a related hypothetical situation: an agent reviewer could approve a technically correct implementation of a feature the team had repeatedly decided against building. Its automated workflows already consult product ownership documents (I'm calling this "git blame for org design"). An agent with that context can retrieve a settled decision or locate an owner. Some of that context is tacit, colloquial, or dangerous to commit to paper. Software can carry an organization’s decisions into the workflow; it cannot confer authority the organization has withheld.

Running the factory requires delegation alongside capability. Engineers and designers need room to investigate, prototype and act within an agreed scope. Companies can also authorize teams to run experiments within limits, keeping decisions close to the work. Cultures can emphasize openness to experimentation alongside respect for domain ownership.

How much freedom, how much ambition, can we let the factory use on its own? How much should we?


At the risk of opening up a very scary can of worms, we have to talk about the state of software factories alongside the state of cybersecurity, which feels at a system level like it is a few days from total collapse. On one hand we have Outlook going down, allegedly because engineers who knew how they might fix an issue in such a gigantic system took voluntary retirement inspired by AI productivity targets. On the other hand, it's not hard to imagine adversarial software factories, designed by state and non-state hackers to bring down/compromise critical infrastructure. I'd assume that the bad guys are giving their factories more permission than the good guys.