Designing
agentic work.

I distinguish deterministic rules implemented in software from decisions inferred by an LLM. Permissions, allowed tools and required fields can follow explicit checks. Interpreting a source or deciding what to investigate next may call for model-based judgment.

The question is where flexibility helps and where a decision must obey a defined rule. A prompt can describe a boundary; software must enforce it when it is a hard requirement. Repeatable model output does not, by itself, turn its interpretation into a verified control. I want that distinction to remain visible in the system design.

In expert domains, I keep asking how vague or precise an agent’s instructions should be. “Review this assessment” leaves the quality criteria open. Naming the criteria, required evidence and expected handover makes the task more concrete without prescribing every step.

Instruction volume is part of the challenge. More instructions can introduce assumptions or conflicts; too few leave the agent guessing. I try to make the goal and non-negotiable constraints precise, while leaving room to explore the method. Finding that balance is an ongoing design question.

For GPT-6 Astra and GPT-6.1 Sol, I refer to OpenAI’s shared GPT-6 prompting guidance. The Astra-specific guidance also explores how clear goals, constraints and completion criteria can replace rigid procedural recipes. I test these recommendations against the model and task.

I want autonomy to match the task’s consequences. Ambiguity, contradictory evidence and critical decisions need explicit escalation criteria. Some events only need an update; others should pause the task for human review. The handover needs to explain the uncertainty and the decision required.

Autonomy needs a clear handover as well as permission to continue. I want to know what was completed, what remains uncertain and which decision needs my attention. Defining those boundaries is part of the work of designing an agent.