Break a workflow into single-purpose steps an agent can run.
Put approval gates where money or reputation is at stake.
Bids2Bank runs proposal work as six stages in a fixed order: intake, triage, capture, generation, compliance and submission 1. Each stage reads only the stages it needs; generation, for example, reads capture and triage 2. Every pod seals its whole output as a baton 3.
Design rule: one step, one job, one output you can check. If you cannot say what a step's output should look like, or how you would know it is wrong, the step is too big to hand to an agent.
Bids2Bank's capture strategist shows how a model step can fail safely: when its model call fails, it logs a degraded event and a deterministic fallback runs 4.
Check 1
Where a person must approve
In Bids2Bank, a proposal package is marked submitted only after a person types the approval phrase 5. An optional review checkpoint can only be continued by a signed-in person, not an API key 6.
Supplier outreach follows the same pattern. With the hold setting on, automatic e-mails, texts and calls are recorded for a person to approve instead of sent 7; sends a person starts by hand are never held 8.
Put the approval where the action cannot be taken back: money leaves, a message reaches someone outside the company, or a commitment is made in the company's name. Let the agent draft and queue everything before that point.
Check 2
Limits an agent cannot talk its way past
Instructions hidden in content an AI system reads can redirect it 9.
So do not rely on a prompt alone to hold a limit. Put caps, holds and forbidden recipients in code the model cannot change.
Bids2Bank's automatic e-mails stop at the lower of the workspace's own limit and the plan's daily cap 10. A workspace with too many complaints or bounces in 30 days has its automation paused until they age out 11. A send paid with prepaid credits is charged before it goes out 12, and the supplier automation sends nothing to DLA 13.
Worked example. A team builds an agent that drafts follow-up e-mails to suppliers. For the first month every automatic send waits for a buyer's approval, automatic sends are capped at 20 a day in code, and no message may go to a government address.