Written for the Director of Restaurant Technology — the seat that owns the handhelds and the store systems, inherits every integration, and has learned the hard way that every rollout inherits every supplier's format forever.

The tax has a precise shape. Your supplier set spans a decade of integration vintages: the distributor whose systems still emit EPCIS 1.1 XML from a 2016 project, the produce network on CSV, the protein supplier whose new SaaS exports its own JSON. Every chain-wide initiative — RFID receiving, FSMA 204 readiness, the next DC expansion — begins by paying the same toll: get sixty-some feeds into one shape, again, by hand, with the institutional knowledge living in whoever did it last time.

The claim of this brief is that the toll booth is exactly the wrong place for humans to stand, and exactly the right place to seat an agent.

Why translation and validation are agent-safe

Delegating work to an agent is a risk decision, and the risk profile here is unusually clean, for reasons that are properties of the work:

The verbs are pure. Translate takes a document in one dialect and returns it in another. Validate takes an event and returns a verdict against the pinned official GS1 schema. Neither mutates anything, neither needs judgment, and both are exposed as read-only-hinted verbs an agent can fire inside already-granted scope without asking anyone mid-task.

Every job is verifiable. A translation is not trusted; it is checked. Each job produces a round-trip fidelity report — what went in, what came out, and whether the round trip is faithful — so the agent's work arrives graded, not asserted. The machinery is the developer side of this family: the translate, validate, and hash doors at epcis.dev answer over HTTPS against sha256-pinned official artifacts, and an engineer on your team can probe them without a meeting.

Failure is loud and typed. A feed that does not validate is refused with the standard's own exception types, not a vendor error string. Refusals queue for a human; they never silently pass.

The worked run

One onboarding sweep across a supplier set of the size a multi-unit chain actually carries — illustrative numbers, realistic mix:

supplier-onboarding sweep · seat: integration-04 sponsor: restaurant-tech team acct · mandate: translate + validate only feeds in .................... 63 EPCIS 1.1/1.2 XML .......... 41 CSV (per-supplier layouts) . 14 vendor JSON ................. 8 translated + validated ...... 57 per-supplier fidelity report attached clean round-trip ........... 49 translated with warnings .... 8 (deprecated vocab, mapped + noted) human adjudication queue ..... 6 ambiguous lot-grain ......... 3 identifier checksum fails ... 2 undeclared custom fields .... 1

The proportions are the argument. Fifty-seven of sixty-three feeds never needed a human, and the six that did were genuine judgment calls — a lot-grain ambiguity only someone who knows the supplier can rule on, an identifier that fails its checksum and needs a phone call. Volume through the agent, judgment through the human, and the boundary between them drawn by the work itself rather than by anyone's patience.

Note what the agent is not authorized to do. The mandate is translate and validate — nothing that writes to the record, nothing that grants, nothing that contacts a supplier. That is deliberate mandate design: onboarding is high-volume, verifiable, low-authority work, and the seat is scoped to exactly that. The pattern — ceiling, scope, distinct observer identity on every act — is what a seat is, and every job in the sweep above is recorded with the agent as the attested observer and the sponsoring team account as the warrantor behind it. Two grains, never collapsed.

The fidelity report is the trust instrument

The reason this delegation works organizationally — the reason your team can accept an agent's output where it would never accept a contractor's zip file — is that the trust never rests on the agent. It rests on the report. Per supplier, per job: what was translated, what mapped cleanly, what carried warnings, what the round trip preserved. Your engineers review evidence, not effort. And the report outlives the sweep: when supplier 44's feed drifts next spring, the delta against its own onboarding report is the diagnosis.

This is also the honest contrast with the field. The developer-EDI vendors who might have owned this problem have left it — the market-structure story is its own brief — and the managed networks answer N-formats with "make your suppliers join." The agent-worked sweep answers it without asking sixty suppliers to change anything, which is the only version of the answer your supplier relations team will thank you for. The chain-side executive read this feeds is RFID and food traceability for restaurant operators.

The mandate design, as a checklist

For the seat you would actually provision:

  1. Authorized: translate any supplier feed; validate against the pinned schema; attach fidelity reports; queue mismatches.
  2. Not authorized: capture to the record; contact suppliers; touch anything outside the onboarding scope. Asks above the ceiling fail closed and land on your team — never a silent grant.
  3. Measured by: feeds cleared per adjudication, and the completeness of the attribution trail.

The toll booth does not disappear. It just stops being staffed by your best people.

If your supplier count is the number that decides every project's timeline, start the interview — it asks about your feed reality, and it ends with exactly one promise about contact.