AI writes code fast. That's the easy part. The hard part is that fast isn't the same as aligned.
Give the same model the same prompt three times and you'll get three different answers. Multiply that across a team of developers — each with their own way of working — and the drift compounds. You ship a lot of code, but very little of it agrees with the rest.
That's the problem Julius opened with in his Usable Fragments talk, "From Fast to Aligned."
Every session starts from zero
The root cause is simple: every AI session starts from a blank slate. The model has no memory of how your team builds. You can write all the standards you want into markdown files, but they land in different spots in the context window — and what gets lost varies by model and by run. Start ten conversations and you get ten blank slates.
Usable as the grounding layer
So the team built an alignment stack with Usable as the grounding layer — a single source of organizational memory that every session starts from. Instead of beginning from nothing, each session begins with the team's accumulated standards, conventions, and hard-won lessons.
The loop looks like this:
The clever part is the last step. When a feature is done, the agent doesn't just open a pull request — it commits what it learned back to Usable: the decisions, the dead-ends, the path it actually took. The next time anyone (or any agent) enters plan mode, that knowledge is already there. One developer's fix instantly becomes everyone's.
Plan time, not review time
The payoff shows up in review. Instead of nitpicking "wrong package manager," "wrong linting framework," "that's not how we structure this," the boring stuff is already handled by the grounding layer. Reviews drop from an hour to a few minutes, and the time you do spend goes to functionality, not formatting. You move the effort to plan time instead of review time. Before, the team was the bottleneck; now they can keep pace with the AI.
It compounds — and it's not locked in
Because every loop writes back, the system gets better the more you use it. A bug in production kicks off an investigation loop that documents the root cause back into memory, so the next encounter is faster. And because the memory layer is LLM-agnostic, none of it is hostage to a single vendor — switch from OpenAI to Anthropic for a better deal and keep every bit of institutional knowledge. (That was the part of the room that got the biggest reaction.)
Watch the full talk
Julius walks through the whole alignment stack, plus a live demo querying benchmarks across 130,000+ fragments in Usable Chat and building a PRD from a conversation on his phone. It's the first of four talks from Usable Fragments 2026.
Usable is knowledge management for AI agents. Connect your LLM, capture your team's knowledge, and stop starting from scratch → usable.dev
Watch the full series → Usable Fragments 2026 on YouTube