Unit 10.00: Choosing a workflow that justifies a crew

The capstone begins with the decision this course has spent nine modules making harder: does this workflow justify a crew at all?

Four checks on a candidate

Volume, steps that genuinely need judgement, steps with writable rules, and agents differing by tools or permissions.

The code runs a supplier-invoice candidate through all four.

import json

candidate = {
    "workflow": "supplier invoice approval",
    "volume_per_month": 900,
    "manual_minutes_each": 6.0,
    "steps_needing_judgement": ["extract from varied PDF layouts",
                                "match line items to contract terms"],
    "steps_with_a_writable_rule": ["amount ceiling", "currency check",
                                   "approver by amount band"],
    "agents_justified_by": {"Extractor": "different tools",
                            "Contract checker": "different knowledge source",
                            "Approver": "no tools, decision only",
                            "Poster": "the only WRITE permission"},
    "why_not_one_agent": "one agent would need write access to post payments",
}
print(json.dumps(candidate, indent=2))

checks = [
    ("volume justifies the build", candidate["volume_per_month"] > 100),
    ("some steps need judgement", len(candidate["steps_needing_judgement"]) > 0),
    ("some steps have rules", len(candidate["steps_with_a_writable_rule"]) > 0),
    ("agents differ by tools or permissions", len(candidate["agents_justified_by"]) > 1),
]
for check, ok in checks:
    print(f"  {'OK  ' if ok else 'FAIL'} {check}")

why_not_one_agent is the field that carries the argument: one agent would need write access to post payments, so it would hold that permission while also doing extraction and judgement.

Note that the candidate lists more steps with writable rules than steps needing judgement. That is a healthy ratio - it means most of the workflow lands in the flow where behaviour can be proven rather than measured.

The mistake this prevents

The mistake is choosing a workflow because it is interesting rather than because it clears the four checks. A low-volume workflow with one judgement step is a good afternoon's work and a bad capstone, because nothing about it will exercise the controls you spent this course building.

Takeaway

Choose a workflow that clears all four checks, and write down why one agent would not do. If the answer is about permissions, the design is already half-argued.

Mark complete