Stale-plan detector & killer for parallel coding agents
Problem
Developers now run multiple coding agents (Claude Code, Codex etc.) in parallel and don't watch them closely. When plans change mid-flight, an agent keeps grinding for hours on an abandoned goal and then lands a green-tests patch for work nobody wants anymore — polluting the diff and wasting compute and review time. There is no tool that reconciles the current intent with what agents are still working on; everyone cancels manually or discovers the waste at final review.
Opportunity
A lightweight supervisor that tracks all running agent sessions against a shared, living task list, detects divergence (goal removed/changed while agent still running), prompts the developer to cancel or redirect, and clearly separates 'run completed' from 'work still wanted'. Timing is ideal: agentic coding exploded but orchestration hygiene tooling hasn't caught up.
Market analysis
Multi-agent managers like claude-squad and Conductor already orchestrate parallel coding agents, but none reconcile running sessions against a living task list; the stale-intent detector is a genuine gap. The risk is distribution rather than feasibility: incumbent supervisors are free and one feature away from this.
Market · Heavy users of agentic coding tools (Claude Code, Codex, Cursor) running several sessions in parallel; fast-growing segment with early-adopter culture and active community discussion.
Pricing · Direct comparables are mostly free and open-source (claude-squad) or cloud services with bring-your-own-keys (Conductor), which caps standalone pricing; a realistic wedge is a small per-seat subscription bundled with broader agent-hygiene features.
Pros
- + Pain is fresh, specific and increasingly common as parallel agents go mainstream.
- + Narrow, buildable MVP: watch sessions, diff against the task list, prompt to cancel.
- + The 'completed' vs 'still wanted' distinction is a memorable editorial framing.
Cons
- − Claude Code's built-in agent teams and free tools like claude-squad can absorb the feature quickly.
- − Detecting that intent changed is fuzzy heuristics, not a deterministic check.
- − Audience of power users, many of whom script their own fixes.
Existing / similar tools
Source
Hacker News (Ask HN)
The technical core is intent reconciliation: diffing task lists, plan files and session transcripts to decide that a goal no longer exists. That is a taste problem rather than an infrastructure problem, which is exactly where a solo builder can beat a platform team. The moat is close to zero, so speed and becoming the default recommendation in the HN and Reddit threads where this pain is described matter more than feature depth. One non-obvious angle: the real cost is not wasted compute but polluted diffs and review queues, and framing the product around clean, intentional diffs will resonate more than token savings.