Our thesis
Coding is becoming free. Taste is the new bottleneck.
With Claude Code, Codex and a wave of software factories, building software has never been easier and it's only getting easier. As the cost of writing code falls toward zero, the hard part is no longer how to build. It's deciding what to build and in what order. We call that taste.
Coding was never the bottleneck
The software development lifecycle is far bigger than typing code. Discovery, prioritization, design, specs, review, QA, release, feedback and iteration. Writing the code was always a thin slice of the work.
For decades that slice was the constraint, so the entire industry optimized it: faster editors, better languages, more libraries. Now agents are dissolving it. When code is cheap, the value moves to the parts that were always hard for machines: judgment, priorities and knowing which problem is worth solving.
The right way to understand this shift is the way we understand self-driving cars. Autonomy doesn't arrive all at once. It climbs in levels. Each one hands more of the doing to the machine, until the only thing left for a human to supply is intent: where do we want to go?
The autonomy ladder
Five levels of autonomy in the SDLC
Borrowed from the SAE's framework for self-driving cars, from driver assistance to full autonomy and mapped onto how software gets built.
- Level 1
Assisted coding
Cruise control and lane-keeping. The human drives, the machine smooths the work.
Autocomplete, tab completion and in-editor assistance. The engineer is still doing every line, but the friction between thought and keystroke collapses.
Unlocks: Massive individual productivity gains.
Cursor
Took the market with AI-native editing and autocomplete.
GitHub Copilot
Brought inline completion to the mainstream.
- Level 2
Autonomous CLI agents
Hands off the wheel on the highway. You supervise closely, ready to take over.
Terminal-native agents take a task and execute multi-step changes on their own. Spec-driven development and TDD become the steering wheel. You describe intent and constraints, the agent writes and iterates.
Unlocks: Whole tasks, not keystrokes, get delegated.
Claude Code
Agentic coding from the terminal, end to end.
OpenAI Codex
CLI agents that plan, edit and run.
- Level 3
Software factories
Robotaxi in a mapped city. Assign a destination, get dropped off, review the ride.
Coding agents become teammates on the board. Assign a ticket, get back a pull request. Work runs in cloud sandboxes; humans triage, steer and review diffs instead of writing the code themselves.
Unlocks: Kanban-style delivery: issues in, reviewed PRs out.
Factory.ai (Droids)
"Software development is more than just coding." Droids across the SDLC.
Linear Agents
Delegate an issue, get a diff from a cloud coding session.
Ramp
Internal agent fleets inspecting and shipping work at scale.
- Level 4
The closed feedback loop
Where Darin livesFully autonomous in its domain. The car decides the route from where you need to be.
The system listens. It captures user feedback, product analytics, customer conversations and bug reports, turns them into prioritized tickets, ships the fix and closes the loop, measuring whether the change actually worked.
Unlocks: Software that observes itself and improves without a prompt.
Darin
The decision layer: signal in, prioritized specs out. This is our home base.
- Level 5
Self-driving software
Set a destination and step out. The car keeps going until you arrive.
You give the software a goal. It keeps working, learning from users, choosing what to build next, shipping and checking progress, until the problems under that goal are solved, or you decide to stop.
Unlocks: Name the goal once. The software does the rest.
Where Darin is headed
Decision memory today. Self-driving product work tomorrow.
Darin is the navigator, not the engine
Levels 1 through 3 automate the doing. The factories can write the code, open the PRs and ship. But a factory pointed in the wrong direction just produces the wrong thing faster.
Darin automates the deciding. We read everything your customers and team are telling you across Slack, support, calls and analytics, then hand back clear priorities and specs, with the reasoning behind every call. That's the taste layer the rest of the stack is missing.
Today that makes Darin the decision engine that feeds the software factories. Tomorrow, it's the closed feedback loop of Level 4: capturing signal, deciding what matters, dispatching the work and measuring whether it landed, all without waiting to be asked.
Level 5 is the full version: self-driving software. You name the goal. The system keeps going until the work is done.
Help us build the taste layer
If code is free, the teams who win are the ones who decide best. Join the waitlist for early access to knowledge that keeps those decisions current for your team and your agents.