Inspiration
AI agents are effective at individual coding tasks, but long research and engineering workflows are still fragile. A failed command, closed session, exhausted context window, or interrupted background process can leave the user repeatedly deciding what command to run next.
I encountered this problem while running long materials-science workflows involving notebooks, simulations, data processing, and repeated validation. I wanted the user to define the goal and safety boundaries once, then intervene only when genuine judgment was required.
What it does
Autopilot NodeKit turns a project request into a durable, verifiable, and auditable task graph.
Each task has dependencies, success criteria, selected memory, a verifier, and persistent execution evidence. Background Codex workers execute ready tasks, while an operator handles routine recovery actions such as creating focused repair tasks, resolving successfully repaired failures, and recovering stale runs.
Human gates remain mandatory for dangerous operations, expensive computation, credentials, destructive changes, and final approval.
How I built it
I built NodeKit as a Python CLI backed by SQLite and YAML project specifications. It includes smart project initialization, task-graph scheduling, non-lossy structured memory, cross-platform background execution, verifier-authoritative completion, shell-safety checks, repair workflows, audit events, and regression tests.
OpenAI Codex with GPT-5.6 was used as the primary implementation and debugging environment. GPT-5.6 helped convert failures observed in real long-running workflows into control-plane rules, implement recovery and scheduling logic, and develop tests and documentation.
Challenges and lessons learned
The hardest problems were distinguishing a slow task from an abandoned task, preventing old repair branches from blocking current work, preserving useful context without destructive summarization, and preventing verification commands from accidentally performing real operations.
The main lesson was that reliable agent autonomy requires more than a longer prompt. It requires explicit state, deterministic verification, evidence-preserving memory, bounded recovery, and clear human authority.
Log in or sign up for Devpost to join the conversation.