Why Your Agent Can't Follow a Plan (And How to Fix It)

You give an agent a complex goal. It starts well, then halfway through it forgets what it was doing, repeats work it already completed, or gets stuck when one step fails and blocks everything downstream. The LLM isn't the problem. The workflow architecture is. I've been building production agents for a while now, and the same three failure modes come up every time: Implicit task structure — the agent doesn't have an explicit list of what needs to happen and in what order No failure isolation — when step 7 fails, steps 8, 9, and 10 all get blocked unnecessarily No resumability — if the process crashes at step 14 of 20, you start over from step 1 Here's the architecture I now use for any workflow that's more than 3 steps. The Core Abstraction: TaskTree Instead of letting the agent free-form plan in its own context window, I make the plan explicit and executable : @dataclass class Task : id : str = field ( default_factory = lambda : str ( uuid . uuid4 ())[: 8 ]) name : str = "" descriptio