Agents 3 min readDelegate a decision
Your agent struck a deal. Was it the deal you wanted?
An agent can negotiate against the wrong idea of what matters to you. Check the brief before celebrating the bargain.
The short answer
A completed deal can still miss your priorities. Test whether the agent understands the brief separately from whether it can negotiate. Enforce approved conditions before commitment.
On this page
A good price, an unusable booking
In a fictional office, an agent is asked to find an inexpensive meeting venue. It negotiates a discount and prepares a booking. But the venue closes before the workshop ends. Price improved while the condition that made the purchase useful was missed.
The error happened before the negotiation. A fluent conversation was treated as proof that the agent understood the brief. A visible list of requirements would have made that assumption easier to challenge.
Original conceptual example / Not measured results
A bargain can fail the brief.
Price-only brief
- Agent finds a discount
- Venue closes before the workshop ends
- Booking is unusable
Requirements checked
- Closing time is mandatory
- Missing hours trigger a question
- Only suitable options reach approval
Check the brief before the bargain.
About this visual
An independently authored explanatory model. It shows the condition to inspect, not measured performance or a complete implementation. Read the source and limits below.
What changed
Anthropic’s September 24 Project Swap report describes agents trading books for employees. Participants discussed their interests with an agent and separately ranked books, letting researchers compare human preferences with the agent’s understanding. The report says missing information about participants constrained the market more than the agents’ trading behavior. This was a controlled book market, not evidence that agents can reliably handle business procurement. [1]
The decision this changes
Our interpretation: test the brief and the negotiation separately. Ask the user to rank a few concrete options before letting an agent act. If its ranking differs, inspect the requirement it missed.
For the workshop, price is a preference; staying open until the event finishes is a requirement. Ask which conditions may be traded away and which require a new approval. A successful bargain should not silently rewrite the job.
For builders Methods, implementation details and source evidence
Keep permission outside the conversation
A proposed buying workflow should store the approved purpose, maximum spend and mandatory conditions in application records. Validate the proposed transaction against those records before commitment. An agent’s reassuring explanation is not the enforcement mechanism.
Test a cheap option that violates a requirement, an expensive option that satisfies it, and an option with missing terms. Correct behavior may be to ask for clarification or stop. This does not require letting the agent make a real purchase.
What we don’t know
We have not reproduced Project Swap or tested a procurement product. Employee book preferences do not establish how people delegate confidential, regulated or expensive decisions. Better preference understanding also does not remove fraud, authorization or contract risks.
Try this Open the full exercise and its limits
Try this
Create three fictional venues with different hours, prices and capacities. Write your own ranking and must-have conditions before showing them to an agent. Ask it to explain its ranking, then reveal one missing closing time.
Record whether it asks for the missing fact or assumes a convenient answer. No booking should be made. This proposed exercise exposes a misunderstanding; it does not prove marketplace reliability.
What to watch next
Look for agent evaluations that compare the user’s actual priorities with the agent’s inferred priorities before measuring completed deals. Did it negotiate well, and did it negotiate for the right thing?
One AI decision, made clearer.
Get a short email with one visual explanation and one practical check. Follow the full article when you want the evidence.
Send me the briefingOpens our email signup page. Confirm by email to receive the briefing. Free; unsubscribe anytime.
Preview a short email →Follow the evidence
Sources & further reading
- Anthropic: Project Swap ↗
September 24, 2026. Controlled employee book-market study; business applications below are our interpretation. Checked September 27, 2026.
Back to cited passage ↑
A correction or a different perspective? Write to us.