🤝 Top Collaborators
🤓 Latest Submissions
TeamWeave
The problem AI agents are quietly becoming coworkers in company after company — and each one is somebody's side project. Marketing runs a summarizer on a cron job, engineering wires copilots into CI, an analyst keeps a GPT script under their desk. These shadow automations multiply, and three failures follow. First, no shared picture. Nobody — not the teams, not their managers, not security — can see which agents exist, what data they touch, or what they cost. Second, no governance. An agent's output flows into another team's work by paste-and-pray: no contract for what crosses the boundary, no review before its results are used, no way to send work back. Third, no provenance. When a downstream result is wrong, nobody can reconstruct which agent produced it, what it was told, or what it actually changed. Agent output gets trusted blind because auditing it after the fact is practically impossible. The solution We built TeamWeave, a canvas for governed agent teamwork shaped like the org itself. It targets the people who own agent workflows — platform and engineering leads, team leads, and the security/reviewers they answer to. Users authenticate into organizations, teams, and roles; each team owns a board — a space on the shared canvas — where they build directed graphs of agent nodes. The contract is the core primitive: every edge between nodes is an explicit JSON edge contract, so a node receives its upstream's summary, results, commands, and artifacts as structured input — not pasted text. A dependency runner walks each graph and pauses on gate nodes: dependent work continues only after a human approves, and reviewers can send it back for rework. Teams hand off instead of throwing over walls. A node sent to another team appears there as an inbox node carrying full provenance — the source team and node travel with the work. Every execution is a frozen attempt, persisted with its exact prompt, executor configuration, and output. Workspace changes are captured as Git snapshots, so what an agent changed is a diff you can actually read. Nodes execute through real IBM Bob runs under Bobcoin caps, or through pluggable model providers (OpenAI-compatible endpoints, OpenRouter, z.ai) configured per node, with a mock provider for zero-cost dry runs. The baseline runs on a local Fastify + SQLite backend. Why now Agent sprawl is new: organizations are crossing a dozen production agents only in the last two years, and the failure modes above are surfacing now. Canvases like Miro and Figma orchestrate people; dev tools orchestrate agents inside one repository. Neither treats agents as managed workers with budgets, contracts, and audit trails spanning teams. TeamWeave does: an agent is a managed worker — budget-capped, contract-bound, reviewed before dependent work continues, attributable after the fact. Demo: https://anton15k.github.io/IBM_HACK/ — project overview with the local run recipe; the full product (Fastify backend + Vite UI) runs from the repo on desktop-class browsers.
27 Sep 2026
