Engineering, woven.
Software development for the agent era. A multiplayer workspace where teams shape intent, agents execute, and review checks the artifact against the plan — not every diff.
Teams shipping with Weaver
The pipeline collapsed.
Software development was built around a slow, expensive step: writing code. Tools, processes, ceremonies — all of it designed to manage that constraint carefully. Story points. Sprint planning. Pull request review. The whole pipeline existed because implementation took days.
Agents collapsed that. Implementation now takes minutes. The work didn't disappear — it piled up at the end. Alignment, planning, and judgment now happen at PR review, where reviewers second-guess decisions that should have been made before the first line landed.
Weaver moves the work back to where it belongs. Teams shape intent together, in prose. Agents execute inside that intent. Review checks that the artifact matches the plan — not that the code is acceptable.
Where Weaver sits
The empty quadrant.
Coding agents excel at execution but live in your editor; your team doesn't. Project tools are built for teams but can't run your code. Weaver is the one place that's both.
player
Lonely planning tools. Hard to find — most builders skipped this quadrant entirely.
Cursor · Claude Code
Code agents that fly inside your editor. Fast, capable, alone. Your teammates can't see the prompt history, the plan, or the diff forming in real time.
player
Linear · Jira · GitHub
The team is here. Plans get written, work gets tracked, reviews happen — but the agent isn't in the room. Code lives somewhere else and gets handed back over the wall.
Weaver
Multiplayer specs the team writes together. Agents executing inside that spec, in front of everyone. Review of the artifact against the intent, not the diff. Both halves of the loop, one surface.
Steered through the chat
Humans decide where it matters.
Agents fly through the easy parts. At every destructive moment — deleting files, running shell commands, opening PRs — Weaver pauses and asks. One click resumes the run.
The structural permission gate runs before the destructive tool call, not after. A prompt-injected agent calling shell directly still round-trips through the prompt first.
Reading apps/api/alembic/versions/0042_add_user_index.py …
The migration is half-applied on staging. Rolling back manually before a fresh re-apply.
Agent · permission request
About to run
rm migrations/0042_add_user_index.py
in the worktree.
Specs in practice
Specs are the prose your team already writes.
Not yet-another-DSL. A node is a paragraph or three. The agent reads it, asks the questions that matter, and waits for you before it ships.
Add a soft-delete column to comments and keep the existing read paths working.
Agent · clarifying question
Should comment threads with all soft-deleted children still appear in the comment count? (Default: no — match the existing thread visibility rule.)
Migrate the rate-limit store from in-process to Redis. Keep the current rate-limit semantics.
Agent · clarifying question
I'll keep the per-user + per-IP buckets. One question: should we fail-open if Redis is unreachable? (Recommendation: yes, with a 1-minute alert threshold.)
Wire the design-partner application form on the marketing site to an api endpoint that emails the team.
Agent · clarifying question
Email-only is the v0; want me to also create a Linear issue in the design-partner project? That keeps the queue with the rest of intake.
Local + cloud
One contract. Two substrates.
Run the agent on your Mac with full access to your repos and shell, or in a cloud sandbox when the team scales. Same loop, same review surface, same permission gates — only the substrate changes.
Desktop
Native macOS app. The agent runs on your machine; credentials and repos never leave it. Permission prompts on every destructive call.
Cloud
Isolated sandbox per run. Same agent loop, same review surface, same permission gates — only the substrate changes. Coming next.