Rivet Whitepaper

Rivet turns reusable engineering knowledge into verified coding-agent workflows.

The paper starts from a simple problem: capable AI agents are still being driven by messy, one-off prompts. A model can inspect a repo, edit files, run commands, and continue across multiple turns, but the process around that work is usually implicit. Architecture gets re-decided from scratch, important constraints drift across chats, validation is inconsistent, and failed work is often repaired with more prompting instead of a better system.

Rivet is my answer to that gap: a development harness that turns engineering judgment into reusable workflow structure. The goal is not to build another model or another giant prompt. The goal is to capture how good work actually happens - standards, skills, templates, references, evals, approval gates, and recovery paths - then make that method executable by coding agents.

The core idea is graph engineering. A serious workflow should know its current state, what node is running, what decisions unlock the next step, what evidence is required to continue, and where a failure should go for repair. The model can still reason inside each node, but the harness controls the context it receives, the tools it can use, the output it must produce, and the gate that decides whether the work is actually done.

That matters because everyone building with AI will eventually need their own techniques. Engineers need repeatable ways to ship agents, auth, workers, data pipelines, and evals. Operators need repeatable workflows for analysis, reporting, and internal automation. Founders and CEOs need a way to encode how their company should use agents without relying on scattered prompt docs or one person knowing the magic words.

Rivet makes those techniques portable. A team can define a skill for building an OpenAI agent, adding Clerk authentication, or designing a durable worker, and that skill can include instructions, references, scripts, templates, and evals. Over time, the system becomes a library of how the team builds: not just what the agent said once, but the actual process, checks, and repair logic that made the work reliable.

The long-term vision is a workflow layer that compounds with use. A run ledger records questions asked, decisions made, files changed, commands executed, gates passed, failures hit, and improvements proposed. That gives teams a way to turn repeated agent work into an improving engineering system: scoped knowledge, controlled execution, verifiable gates, typed recovery, and a reviewed feedback loop instead of ad hoc automation.