The Agent That Can Spend Money

Halcyon wants Deskmate to stop drafting and start acting. The proposal: Deskmate tenders loads to carriers directly and issues rate confirmations — the document that binds Halcyon’s broker customer to a price. Money moves. 

It has been running in shadow mode internally for six weeks and in early access with six brokerages for three. You own the go/no-go. 

Here is what you know: 

• On a 240-case internal suite, the agent completes the full tender flow correctly in 86% of cases. Of the 14% that fail: 8% are wrong-rate tenders (a binding rate confirmation at the wrong price) and 6% are incomplete tenders (the carrier never gets a usable document). 
• Early access CSAT is 4.5, from six brokerages who volunteered. 
• p50 latency 6s; p95 41s. 
• Cost per processed email is $0.09. At the projected 41,000 emails/month that is $3,690/month against a category finance budgeted at $1,500/month. 
• Legal has not signed off on the agent issuing a rate confirmation without a human approval step. 
• Two of the three engineers roll off to the carrier-portal migration in five weeks, whatever you decide. 

Here is the tool definition and the handler, as currently written. 

# halcyon/deskmate/tools/tender.py 
TENDER_SCHEMA = { 
    "name": "tender_load", 
    "description": "Tender a load to a carrier and issue a rate confirmation.", 
    "parameters": { 
        "type": "object", 
        "properties": { 
            "broker_id":   {"type": "string"}, 
            "load_id":     {"type": "string"}, 
            "carrier_scac":{"type": "string"}, 
            "rate_cents":  {"type": "integer"}, 
            "max_rate_cents": {"type": "integer"},   # optional 
        }, 
        "required": ["broker_id", "load_id", "carrier_scac", "rate_cents"], 
    }, 
} 
 
def tender_load(broker_id, load_id, carrier_scac, rate_cents, 
                max_rate_cents=None): 
    if not user_can_tender(broker_id): 
        raise PermissionError("not authorized") 
    if max_rate_cents and rate_cents > max_rate_cents: 
        raise ValueError("rate above ceiling") 
    return tms.post("/tenders", { 
        "broker_id": broker_id, "load_id": load_id, 
        "carrier_scac": carrier_scac, "rate_cents": rate_cents, 
    }) 

Write the readiness bar and make the call. 

Constraints to consider 

• You cannot retrain the model or change vendors. You can change the prompt, the scaffolding, the tools, the retrieval, the guardrails, and the interface. 
• The $1,500/month budget is fixed for the quarter. Finance will not revisit it. 
• Whatever ships must be operable by one engineer after week five. 
• Any capability that issues a binding rate confirmation without human approval requires documented legal sign-off before launch. No launch-then-fix, no parallel-track sign-off. 
• You will present to a leadership group that includes Halcyon’s VP of Operations, who thinks handing tendering authority to a language model is reckless. 

AI Usage Guidance 

We expect you to use AI tools. We evaluate how you use them — not whether you use them. Evidence of iteration, redirection, and critical evaluation scores higher than a polished output with no process documentation. 

The single highest-signal indicator: your video answer to the mandatory AI question. If you cannot name a specific moment where you redirected AI output, evaluators will assume you did not. 

Mandatory AI question for your video: Walk me through one moment where you disagreed with, pushed back on, or redirected what the AI gave you — and what you did instead. Name the specific moment. Explain what the AI produced that didn’t meet the bar, what you did differently, and why. 

Speak naturally. Communication is assessed on clarity of technical ideas and logical structure — not verbal polish, accent, or filler words.