Manifesto
The bottleneck has moved.
When agents write code in minutes, the cost of building shifts to deciding what to build. Specs are the new atomic unit. Humans steer where it matters. The rest belongs to the agent.
1. Writing code is no longer the bottleneck.
Five years ago, the most expensive seat at the table was the one writing code. The constraint was implementation: how long until this lands, how risky is the change, how many engineers does it take. Tools and processes grew up around that constraint — Jira tickets, sprint planning, story points, PR reviews. They optimized for the bottleneck of the day.
The bottleneck has moved. Agents now produce working code in minutes for problems that used to take days. The cost frontier shifts up the stack — from how to build it to should we build it, to did we build the right thing. The expensive seat is no longer the one writing code. It's the one deciding what to write and judging whether what came back is right.
2. The node, not the task.
A task is a card on a board. A node is the intent itself — written in prose, edited collaboratively, persistent across the lifecycle of the work. Agents read nodes; humans shape them; tasks emerge as a side-effect, not a primary surface.
Why this matters: prose carries nuance that fields don't. "Add a soft-delete column to comments and keep the existing read paths working" is a complete spec. Translating that into "Title", "Description", "Acceptance criteria", "Priority", "Assignee" loses information. Worse, it pushes the alignment work back onto humans at the moment they're least equipped — at PR review, after the agent has already shipped.
3. Steered through the chat.
Free execution between natural stops. The agent flies through file reads, refactors, test runs. At every destructive moment — running shell commands, opening PRs, deleting files — Weaver pauses and asks. A click resumes. A click denies.
The permission gate is structural, not decorative. It runs before the destructive tool call, in the toolset itself. A prompt-injected agent that calls shell directly still rounds-trips through the prompt. There's no "trust mode" that bypasses it.
4. Intent-level review.
Code review optimized for the era when humans wrote every line. Reviewers traced control flow, caught off-by-one errors, second-guessed naming. That's a poor use of human attention when the agent already ran the tests, opened the PR with a clean diff, and passed CI.
Intent-level review asks a different question: did we build the right thing. The reviewer reads the node prose, looks at the artifact summary, checks the test results, and either approves the intent or sends it back. The diff is below the fold — available for when it matters, not the primary surface.
5. Local-first, cloud-when-you-want.
The desktop app runs the agent on your Mac. Your repos, your shell, your credentials — Weaver never holds them. When teams scale and run parallel agents, the same agent loop targets a cloud microVM. The sandbox interface is one Protocol; the substrate is a configuration choice.
6. Shape your team's memory.
Past nodes don't go into a graveyard. Decisions, clarifications, rejected approaches — all of it stays queryable for the next agent run. "See also a prior project" becomes a first-class link, not a Slack search. The team's memory compounds; the agents get sharper as the corpus grows.
Weaver is in a design-partner program. We're working with a small number of engineering teams who share the thesis and want to ship with humans steering AI. If that's you, talk to us.