Long-running coding agents fail in boring ways. They disappear for hours, burn past the budget, call work done without checking the requirements, or hang forever on a bad provider stream. Atomic 0.9.14 and 0.9.15 put those jobs under runtime control. Run it without babysitting it. Workflows now send configurable heartbeats to the parent chat. You can see that a run is alive, inspect its status, and steer it before a missed constraint spreads. Set a budget before the run starts. Give each workflow a budget for elapsed time, charged tokens, and provider-reported cost. Atomic warns the parent chat as the run approaches its limit. If the budget is exhausted, an active stage gets one wrap-up turn before the run stops resumably at an execution boundary. Verify every requirement, not just the tests. Verification now scores each criterion-verifier pair from 1 to 20, requires a passing mean with no veto, re-asks invalid reports, and writes an auditable comparison ledger. A model saying “done” is not proof. Stop provider failures from wedging the job. Stalled streams now settle as retryable transport errors. Workflows can retry or fail over instead of sitting in “running” forever. We also made it easier to keep the transcript visible while answering workflow questions, choose the right skill when names collide, and return typed workflow data without replacing the human-readable artifact. Stay in the main parent chat while workflows run as graphs of coding agent sessions in the background. Attach to sessions, steer, and remain in control as the engineer. Thank you to our contributors and issue reporters for building Atomic with us for these releases: morgan-coded Lingwcy Shreyas Devadiga @makgunay @RayFernando1337 tyler-utah Release notes: