Background Tasks: The One Actor in the Codebase and the SIGTERM Bug That Only Broke on Linux

Our agent can plan multi-step work with a persistent task DAG, compress its own memory, delegate to subagents, and load skills on demand — all driven by the same agent loop from the first guide. But every tool call still blocks. When the model calls bash to run a test suite that takes two minutes, the loop sits idle, waiting for the process to finish before it can do anything else. If someone asks "run the tests and while that's going, create the config file," the agent does them sequentially —