agentic.rayuduramisetti.com

10 roles

Jobs of the future

When the buyer is a program, the work moves. These are the roles that follow from the specs and the failures on the rest of this site, each tied to the job it grows out of.

A caveat worth reading

There is no hiring data for most of these yet. They are not predictions about titles; they are the work that has to happen once software does the buying, grouped the way someone would actually staff it. Where a role is already anchored to a spec, a rule or a live posting, it is linked.

The closest thing to a number: Mentions of the agentic AI skill cluster in US job postings went from 0.06% in 2024 to 0.23% in 2025, about 90,000 postings. Lightcast, for the Stanford AI Index 2026 ↗ It counts a skill cluster, not these roles.

02Commerce protocol integrator

Checkout

Implements and keeps up with ACP, UCP and the MCP bindings on the merchant side, and runs the conformance suites when versions move.

Grows out of
Payments and API integration work.
Why it exists
Two open checkout specs, both versioned, both moving. Someone has to own the merchant's side of them.

04Mandate and dispute analyst

Payment

Reconstructs, months later, what a person actually authorised: reads the signed mandates, decides whether the purchase was within them, and answers the chargeback.

Grows out of
Chargeback and payment operations.
Why it exists
When software does the buying, the dispute question changes from who used the card to what the human agreed to.

05Agent red teamer

Trust

Attacks your own buying flow: injected product pages, poisoned tool servers, replayed payments, carts tampered with between quote and pay.

Grows out of
Application security and penetration testing.
Why it exists
The attack surface moved from the checkout form to everything the agent reads on the way there.

06Product data steward

Discovery

Owns catalogue quality: attributes that are complete, prices as integers, identifiers that survive a re-platform, variants that mean something.

Grows out of
Merchandising and e-commerce operations.
Why it exists
An agent cannot buy what it cannot parse, and a shop with a clean feed beats one with a prettier site.

07Machine pricing manager

Payment

Sets what machines pay: per-request prices, crawl terms, rate limits, and whether a given agent gets a 402 with terms or a 403 with nothing.

Grows out of
Pricing, revenue and API product management.
Why it exists
Agent traffic is either revenue or cost, and today most sites choose cost by refusing it.

08Agent evaluation engineer

Trust

Builds the test set a buying agent has to pass, tunes the thresholds for when it may act alone, and watches them drift once real traffic arrives.

Grows out of
QA, data science and ML evaluation.
Why it exists
An agent that is right most of the time still needs a number that decides when most is not enough.

09Agent policy lead

Trust

Writes the rules before the lawyers have to: spend caps, what an agent may buy unsupervised, what gets logged, and what the customer is told.

Grows out of
Risk, compliance and trust and safety.
Why it exists
Consumer protection law assumes a person clicked. Someone has to decide what the firm does until that catches up.

10Agent incident support

Checkout

Handles the customer whose agent ordered the wrong thing: reads the trace, explains what happened, and gets it reversed.

Grows out of
Customer support and escalations.
Why it exists
The buyer cannot explain itself, so support has to read the transcript instead of asking the customer what they clicked.

What shrinks

The other half of an honest answer.

  • Conversion work aimed at human attention: banners, urgency copy, the cart nudge. Agents do not feel urgency.
  • Keyword SEO as the main discovery channel, where the reader is a model reading a feed instead.
  • Manual order entry and reorder desks, which is exactly the repetitive buying agents do first.

What transfers

Worth more than any one protocol.

  • Reading a spec properly, including the part that says what it does not cover.
  • Designing for a caller with no eyes: errors that machines can act on, states that are explicit.
  • Thinking in probabilities and thresholds, not just pass and fail.
  • Writing the policy down, so a decision can be audited long after the purchase.

Most of these start with the same reading: the protocols for what you would implement, the attacks for what you would defend.