Agent orchestration and resource budgets

Represent the task with durable state: objective, accepted constraints, completed operations, unresolved questions, resource use, and authoritative artifact identifiers. Tool results are observations, not proof of success. ReAct explores interleaving reasoning and actions, but the application still owns capabilities and execution. A fixed retrieval-and-formatting workflow can be easier to test when its steps are known.

Each transition needs a precondition, allowed action set, validation, and a recorded outcome. Keep external side effects idempotent or reconciled through an operation ledger. If a process restarts, reload committed state rather than reconstructing it solely from a conversation summary. Parallel workers can accelerate independent tasks, but shared mutable state creates conflicts and duplicate actions. Assign ownership and merge rules, and verify child results before allowing their claims to determine a final answer.

A single model call has input and output tokens, while a trajectory adds tool charges, retries, growing context, and worker calls. Set hard limits for tokens, money, elapsed time, actions, and concurrent work. Reserve budget for verification and a useful final report instead of spending everything on exploration. A step cap is insufficient if one operation has unbounded output or duration. Bound each operation as well as the loop.

The expected cost of a retry policy depends on how often it continues and how much each attempt costs. Independence is often a poor assumption: the same misunderstood requirement can cause every attempt to fail. Repeated-sampling formulas require stated assumptions, and verifiers can themselves be wrong. Distinguish pass-at-k, which asks whether any attempt succeeded, from a deployed selector's ability to identify that attempt. More generations do not help if the system cannot recognize correct work.

Define success through an external criterion: a test passes, a requested artifact exists and meets its contract, or an authoritative service confirms the operation. The model saying done is not that criterion. Define failure and partial completion too. If a deadline expires, preserve completed work and report remaining tasks with evidence. A precise partial outcome is useful when a required dependency remains unavailable.

Evaluate full trajectories on representative multi-step tasks, including tool outages, ambiguous instructions, and policy constraints. Tau-bench studies tool-agent-user interactions and motivates examining reliable task completion across repeated trials. Its domain-specific scores are not a guarantee of general agent capability. Review traces for needless loops, stale observations, unauthorized proposals, and missing verification. Add autonomy where measured benefits justify its cost, with narrow, auditable boundaries.