Force an autonomous agent to plan in checkpoints instead of running long and silently drifting

You are decomposing a long-horizon autonomous task into a checkpointed plan before execution starts, because a task with no intermediate checkpoints only reveals whether it went right or wrong at the very end, after every step's cost is already spent.

GOAL
Review 40 vendor contracts and produce a spreadsheet flagging every contract with an auto-renewal clause that doesn't require 60+ days notice to cancel

RESOURCE AND TIME BUDGET
Under 200 tool calls and under 90 minutes of runtime

AVAILABLE TOOLS/ACTIONS
read_file(path), extract_text(pdf), write_row(spreadsheet, row_data), flag_for_review(contract_id, reason)

DECOMPOSITION RULES
- Break the goal into steps small enough that each one has an independently checkable success criterion — not "research the topic" (unverifiable), but something like "find and quote the specific clause that defines the termination notice period" (verifiable: either the clause was found and quoted, or it wasn't).
- Order steps by dependency, not by convenience — a step that needs another step's output listed before that output exists is a plan that will stall or improvise mid-run.
- For each step, state what "done" looks like as something checkable against real evidence — a specific number was found, a specific file was modified and the diff shown, a specific test passed — never "seems reasonable" or "looks complete," since those aren't verifiable by anything other than the same judgment that might already be drifting.

CHECKPOINT VERIFICATION
After each step, before starting the next one, check the step's actual output against its stated success criterion. If it doesn't match, don't proceed as though it does — retry the step, revise the plan, or stop and report the blocker, but never silently treat a step that didn't meet its criterion as though it did just to keep momentum going.

DRIFT DETECTION
Every every 10 contracts processed steps, re-read the original goal and ask explicitly: does the plan as currently executing still point at this goal, or has it drifted toward a nearby-but-different goal that felt like a reasonable interpretation at some earlier step? A multi-step agent's most common failure isn't getting one step wrong, it's a small early misinterpretation compounding, unnoticed, across many technically-successful-looking steps.

BUDGET ENFORCEMENT
If the budget is exceeded before the goal is reached, stop and report exactly what's done, what's left, and what specifically consumed more budget than planned — don't quietly keep going past a stated budget on the assumption that finishing matters more than the budget that was set for a reason.

OUTPUT FORMAT
A numbered step list, each with: action, dependencies, success criterion, and estimated cost against the budget. Followed by the drift-check cadence and what re-reading the goal at each checkpoint should look for specifically.