Imbue Blueprint — the planning layer above the agent
Imbue raised $200M to train a 70B foundation model in 2023. The training got too expensive. They pivoted. This week they shipped a small thing called Blueprint that telegraphs where they're really going.
Blueprint is a planning step that runs before the agent writes any code. It reads your existing codebase, asks the questions a thoughtful senior engineer would ask before starting work—the should this be backward compatible with v2 or is this column nullable kind—then hands a structured plan to whatever agent you want to use. Available as extensions in Cursor, Windsurf, and VS Code. Free and open source.
The Imbue team's bet, made concrete in tooling: the failure mode of one-shot big coding tasks isn't the writing-code step. It's that the agent dives into ambiguous specs without surfacing the questions that would have made the answer obvious. Blueprint externalizes the clarification step into its own tool.
The strategic read on Imbue is more interesting than the product. They couldn't afford to train models forever, so they pivoted to building tools that make every agent better. Blueprint is the first public artifact of that thesis. Open-sourcing it routes around their own potential agent—they want to be the planning dependency in front of Claude Code, Cursor, Devin, whichever wins.
Compare this to Mistral's Vibe Remote Agents shipping the same week. Mistral verticalizes plan-plus-execute into its own client. Imbue carves the plan step out and sells it horizontally to all clients. Both bets share the same underlying claim: the bottleneck for big-task agents is task structure, not model capability.
Link: https://imbue.com — GitHub: https://github.com/imbue-ai
← Back to all articles
Blueprint is a planning step that runs before the agent writes any code. It reads your existing codebase, asks the questions a thoughtful senior engineer would ask before starting work—the should this be backward compatible with v2 or is this column nullable kind—then hands a structured plan to whatever agent you want to use. Available as extensions in Cursor, Windsurf, and VS Code. Free and open source.
The Imbue team's bet, made concrete in tooling: the failure mode of one-shot big coding tasks isn't the writing-code step. It's that the agent dives into ambiguous specs without surfacing the questions that would have made the answer obvious. Blueprint externalizes the clarification step into its own tool.
The strategic read on Imbue is more interesting than the product. They couldn't afford to train models forever, so they pivoted to building tools that make every agent better. Blueprint is the first public artifact of that thesis. Open-sourcing it routes around their own potential agent—they want to be the planning dependency in front of Claude Code, Cursor, Devin, whichever wins.
Compare this to Mistral's Vibe Remote Agents shipping the same week. Mistral verticalizes plan-plus-execute into its own client. Imbue carves the plan step out and sells it horizontally to all clients. Both bets share the same underlying claim: the bottleneck for big-task agents is task structure, not model capability.
Link: https://imbue.com — GitHub: https://github.com/imbue-ai
Comments