nerve — One Living Plan for Your Coding Agents

Problem. Agentic coding tools have solved execution but not continuity. Every prompt is treated as a fresh start: change your mind mid-run and you are told to restart the prompt, throwing away everything already in flight; each new prompt re-plans overlapping work from scratch and forgets earlier intent; and whether parallel tasks depend on each other lives only in the model's prose and vanishes when the request ends. Developers pay for the same reasoning twice and lose work they already started. Solution. nerve is a continuity layer that keeps multi-agent coding work in one persistent task graph. A main agent decomposes each prompt into categories with first-class depends_on edges, status and results. When a new prompt arrives, only pending work is morphed: it is re-planned, deduped and re-ordered in place, while running work is immutable and never interrupted. depends_on is validated as a directed acyclic graph, so a category dispatches the moment its dependencies finish, and a failure cascades a blocked state downstream. Execution runs in a detached background pass, so you can keep sending prompts; each one joins the same running plan, and when the pass drains the agent reports results as its own message. Interaction and users. Teams already paying for agentic tools use nerve as a coordination shell around the agents they run. In the CLI (npm install -g nervous-cli, then nerve) they get a terminal-native, two-pane workspace: a live task graph beside the conversation. A browser workspace adds the same engine with a cloud-backed file explorer. They keep steering instead of restarting, and the persistent graph is a readable record of what was asked and why. Why it is unique. nerve is not another executor; it is the memory and scheduling layer agents are missing. Relationships are inspectable data, not prose, so re-steering is safe by design. It closes a structural gap rather than adding a feature, and it is built, not imagined: the persistent graph, immutable running work, dependency scheduling, detached execution and prompt queue all run today, on IBM Granite on watsonx.ai, with the project built through IBM Bob across 84 reviewed pull requests. Efficiency follows the design: a thin orchestrator with no database, and in the CLI your code never leaves your machine.