Knowledge Workers
KW (Knowledge Workbench) is a browser-native OS for actual real world *knowledge work*.
Core trick
Generates a brief and assumptions first, then grades subagent output against a rubric hidden from the orchestrator.
Description
KW (Knowledge Workbench) is a browser-native OS for actual real world *knowledge work*. It’s built for messy, ambiguous tasks where the hard part is framing, tradeoffs, and verification, not just running through a workflow autonomously.
What makes it different from “knowledge work = code
- No single correct output: it produces *options + rationale*, not “the answer.” when in explore mode
- Framing is first-class: it generates a brief and assumptions up front because the problem statement is usually the bug.
- Rubric-driven verification: it grades work against a rubric that the main orchestrator can’t see (prevents self-justifying / goalpost-moving).
- Gap-finding as a feature: it actively surfaces what you didn’t consider - often the real value for managers.
- Checkpointable workflows: you can rerun from a prior state like an investigation trail, not just a chat log.
Workflow (brief → rubric → subagents → verify → revise)
README excerpt
# KW (Knowledge Workbench) A browser-native system for *knowledge work*—not "coding, but with longer prompts." Built for messy, ambiguous tasks where the hard part is framing, tradeoffs, and verification, not compiling. KW uses [kw-sdk](https://github.com/ClioAI/kw-sdk) as its orchestration harness—the engine that handles brief generation, hidden-rubric verification, subagent coordination, and the mode system. Everything else in this repo (the server, frontend, browser sandbox, file system layer, streaming infrastructure) is built around that core. ## Why this exists Most AI tools treat knowledge work like code: one input, one output, done. But real knowledge work—strategy memos, policy analysis, vendor evaluations, research briefs—has no compiler. The hard part isn't generating text. It's figuring out what the right question is, what you're trading off, and whether you actually addressed the thing you set out to address. KW is built around three observations: 1. **The problem statement is usually the bug.** If you jump straight to answers, you optimize for the wrong thing. KW generates a *brief* (scope, assumptions, deliverables) before doing any work—so you can catch framing errors before they compound. 2. **Self-evaluation is broken by default.** An AI grading its own work will move the goalposts to justify what it already produced. KW creates a *rubric* that's hidden from the executor—a separate verifier scores work against criteria the main agent never sees. 3. *