I'm launching Resolv, an AI-powered IDE built on the premise that's increasingly unpopular in the industry: generative models do not think, they only simulate reasoning. They are incredible at exploring possibilities, surfacing context, and amplifying human judgment—but the moment we let them make decisions, we get misalignment, subtle bugs, and architectural drift.
Most AI coding tools today pretend that the model can reason autonomously. They rush to output code, hide tradeoffs, and quietly make hundreds of micro-decisions on your behalf. Resolv does the opposite: it refuses to proceed until every ambiguity, tradeoff, and architectural question is explicitly resolved by you. The workflow is deliberately rigid:
- Supervisor explores your codebase and blocks progress until you resolve every open question or detected inconsistency. - Planner helps you turn intent into a precise, human-approved technical blueprint. - Executor follows that blueprint exactly ("Plan is Law") and surfaces alternatives without ever choosing. - Auditor checks the result against your intent and project standards.
The central interface is the Logic Specification: a structured place where you (not the model) document missing information, and most importantly, all technical tradeoffs and logic decisions before any code is written. It's slower than the "vibe-coding" tools, but it eliminates the expensive hallucinations and technical debt that come from letting models pretend they understand.
It's in early alpha, rough around the edges, but already useful for real projects. Would love feedback from anyone who's frustrated with the current generation of AI agents.
Try it at https://resolv.sh, email me at eric@resolv.sh if you have any questions or comments. Note that there is also a community page on the website.
Thanks, Eric