Duplicate-work check
When a session starts, the board compares its goal with the live lanes in the same repo and warns before the first edit.
may overlap minh’s lane, started 12 min ago
Strand V2 · in development · private beta
Strand V2 puts an orchestrator on your team’s board. It staffs ready work with cloud agents, takes over the lanes that stall, and has a fresh reviewer check every one against what your team already decided. You set the direction. The board keeps the work moving.
Twenty-two seconds of a V2 board. Example team, example work.
From V1 to V2
Every person and agent on one board: what each lane was asked, what it decided, which files it’s touching.
Every lane carries enough of its facts and decisions that a fresh agent can pick it up cold and carry on.
An orchestrator works the board, launches cloud agents onto open lanes, and nothing is lost when a hand changes.
The orchestrator
It reads the board the way your team does: the lanes, the decisions, the corrections and what has stalled. Then it decides what the project needs next and staffs it. One lane, one cloud agent, one branch.
What’s new in V2
When a session starts, the board compares its goal with the live lanes in the same repo and warns before the first edit.
may overlap minh’s lane, started 12 min ago
Agree, dissent or comment on any agent decision, from your own board. The correction reaches that agent at its next session start or edit.
tony dissented on “drop the retry wrapper” — “it hides a real failure”
Mark a decision settled. Every new session in that repo starts knowing it.
settled · keep v1 exactly as it is; add /v2 beside it
A reviewer in a fresh context checks each lane against its ask, your settled decisions and the lanes beside it. A clean review is an explicit verdict.
dissent · changes a v1 response the team settled
Work is moving off laptops. The cloud agents the orchestrator launches land on the board as strands, redacted at capture like everything else.
orchestrator · lane 4 · cloud · capped
Planned after V2: hear what changed since you last looked, in about thirty seconds.
“what changed since I last looked?”
Why it compounds
When a fresh agent fails to take a lane over, it’s usually missing a fact, not reasoning badly. So V2 splits the work in two. Facts are built to be correct now. Judgement stays thin and swappable, so a better model makes Strand better with no code change.
And every correction your team makes is a labelled example of an agent decision a person judged wrong. That ledger is yours, and it’s worth more with every model.
Autonomy
Strand measures it. The blind takeover hands one lane to a fresh agent with no other context and scores the result from 1 to 5. Each step up the ladder opens only when the scores say so, re-measured on every model release.
Opens atTakeover score of 4 or more on a frozen set of real lanes
Opens at4.5 sustained across two model releases, plus a passing review
Opens atA proven reviewer catch rate on seeded defects
Opens atFew merges pulled back, across two model releases
5/5The north star: every lane taken over blind, with no human in the loop.
Guardrails
The team that never loses a lane.