Every team building an AI freight agent rediscovers the same iceberg. The demo — an agent that reads a load and proposes a tender — takes weeks. The production system underneath it takes quarters: workflow state that survives retries, per-customer approval policy, idempotency, audit evidence, integration health, and the dead-letter queue for everything that goes sideways. None of it differentiates your agent. All of it is mandatory.
What the execution layer actually includes
- Load and workflow state: a durable, queryable model of every shipment your agent touches, kept consistent under concurrent updates.
- A policy engine: per-company rules deciding which actions are allowed, approval-gated, or blocked — configurable by the customer, not hardcoded in prompts.
- Approval infrastructure: packets with evidence and reasoning, routing, expiry, and the audit record of every decision.
- Safety plumbing: idempotency keys, duplicate detection, retries, dead letters, and incident records for when providers fail.
- Evidence and audit: every action tied to a company, load, agent identity, and approver — the thing your customer's lawyer asks about.
The trade, honestly
| Build it yourself | Build on Headless Haulbase | |
|---|---|---|
| Time to production | Quarters of infrastructure before the first safe tender | Your agent calls explicit freight actions from day one |
| Engineering focus | Split between your agent and the execution layer | Concentrated on your agent's judgment — your actual moat |
| Safety posture | You design and defend approval and audit from scratch | Policy answers, approval packets, and audit records are the default |
| Customer trust | Earned through your own compliance story | Inherited from an execution layer built for approval-gated freight work |
| Control | Total — including total maintenance | Full control of agent behavior; execution layer is the contract |
When building yourself is right
- The execution layer is your product — you are selling freight infrastructure, not an agent.
- Your freight domain is unusual enough that generic load, tender, and exception models genuinely do not fit.
- You have the team and the quarters, and the opportunity cost is acceptable.
Headless Haulbase exposes the execution layer as a headless TMS API: load context, explicit actions, policy answers, approval packets, and audit history. Your agent keeps its brain; ours keeps the books.