Three Vendors Retired Their Agents. One Named a Date.

Every Sunday I look at what’s moving under the senior analytics craft, and this week it was a deprecation notice with no date on it. On July 30 AWS closed Amazon Bedrock Agents to new customers, renamed it Bedrock Agents Classic, and answered the obvious question in its FAQ: there is no migration deadline, Classic stays available to existing customers in maintenance mode, no planned end-of-life. Then, four questions later, it said the model catalog is frozen as of July 30 and new models will only appear in AgentCore.

Read those together and you have a service that will run your agents forever on a model that stops improving. Nothing breaks. Nothing gets better either. No date for a calendar, no renewal to trigger a review, no invoice that changes. It is the only kind of deprecation that cannot be escalated, because there is nothing to escalate.

OpenAI did the opposite. Microsoft did a third thing. Three notices in a month, none describing the same obligation — which is why the skill repricing right now is not agent-building. It is reading a vendor notice correctly and giving back a plan with dates on it.

What’s actually moving in the market

Three retirements, three different shapes. AWS: no deadline, frozen catalog, and accounts without Bedrock Agents activity in the past twelve months get an AccessDeniedException (HTTP 403) on CreateAgent, with no exception process. OpenAI: the Assistants API shuts down August 26, a hard date published in the migration guide, with Responses as the replacement. Microsoft: no automated or in-place path for Copilot Studio agents built before March 18 — create the new agent, manually reconfigure every channel and connection, decommission the old one. One of these lands in a project plan. The other two land nowhere, which is where unbudgeted work goes to wait.

The economists split, and only one side is on our side. In the first Indeed Labor Market Outlook Survey (Laura Ullrich and Svenja Gudell, August 5), 52% of more than 100 economists expect AI to be at least a mild drag on employment over the next year, 35% a net gain, 13% no effect. Software development landed near the top of both lists, and the two camps named different winners. The economists who expect software to shed jobs named Data and Analytics and IT Infrastructure as the beneficiaries. The ones who expect software to gain named Administrative Support, Marketing, and HR as the losers. Our lane is a winner in one forecast and absent from the other. That is a coin flip with better manners.

The money is chasing the cost problem, not the model problem. Databricks closed $5 billion at a $190 billion valuation on August 13, past a $7 billion run-rate at more than 80% growth, and Ali Ghodsi told CNBC the demand driver is AI cost pressure pushing enterprises toward model routing and open-source options. Read that as a buyer: what is being funded at scale is the layer that tells you what your agents cost and which model answered. That is governance work, and a migration makes you do it from scratch.

My own tracker, with the usual caveat. The JD Salary Tracker digest here last refreshed July 30, so treat these as three-week-old bands: Director of AI Enablement at $139K–$250K, Director of Revenue Analytics and AI Innovation at $196K–$248K, VP of Insights and Data Governance at $150K–$250K. None of those titles is paid to build an agent. All of them are paid to own what happens when the platform underneath one changes.

What I’d do this week

Give the undated notice a date. AWS declined to set one. Set your own.

• User moment: the QBR where someone asks what the announcement costs and nobody has looked.

• Shape: every production workload on a maintenance-mode platform, each with a re-evaluation date and the trigger that moves it sooner — a model you can’t get, a connector with no successor, a renewal.

• Time budget: ninety minutes.

• Artifact: one page, dated, with your name on it.

• Success: the date survives contact with someone senior. The credible failure is finding a workload nobody owns, which is worth the ninety minutes by itself.

Then write the migration one-pager. I have retired two BI platforms — walked one enterprise off Tableau and Sisense onto a consolidated stack — and that work answers the question back in every Director loop this summer: not can you build it, but can you move it without breaking it.

• User moment: “walk me through a migration you ran.”

• Shape: what ran on the platform, what you froze first, what broke anyway, what it cost in weeks and dollars, and the contract clause you’d demand next time.

• Time budget: two to three hours, from memory and old tickets.

• Artifact: one page for a hiring manager or a CFO.

• Success: answering with numbers where most candidates answer with vibes.

One thing to capture before either page: today’s pass rate on fifty real questions your reporting answers. A vendor evaluation of its own production agent (Arunabh Dastidar and the Leni Team, arXiv 2607.17044, July 19) found most of its uplift over the base model came from scaffolding, routing, and specialist models rather than the verification step — the layer a migration rebuilds and nobody measures. It is the vendor grading its own homework, and the limitations section says so. Take the direction, not the decimal.

One counter-read, so the scar tissue doesn’t get overpriced. Every successor platform here is more open than what it replaced, so the migrations may get cheaper and the premium on having survived one may be transitional. The durable asset is the discipline, not the war story. Vendors will keep retiring the middle. The people who can put a date on it will keep getting the calls.

---

Sources

• Amazon Bedrock Agents Classic maintenance mode — AWS documentation, effective 2026-07-30

• Assistants migration guide — OpenAI API documentation, shutdown 2026-08-26

• Recreate Copilot Studio agents with Microsoft Entra Agent ID — Microsoft Learn, 2026-06-15

• The Labor Market Is Tilting Toward Seniority — Indeed Hiring Lab, Felix Aidala & Sneha Puri, 2026-07-23

• Economists Expect a Cooled Labor Market and an AI Reshuffling of White-Collar Work — Indeed Hiring Lab, Laura Ullrich & Svenja Gudell, 2026-08-05

• Databricks wraps $5 billion funding round at $190 billion valuation — CNBC, 2026-08-13

• Databricks Grows >80% YoY, Surpasses $7B Revenue Run-Rate — Databricks press release, 2026-08-13

• JD Salary Tracker — Golden Data, digest refreshed 2026-07-30

• Where Does Agent Reliability Come From? — Arunabh Dastidar and the Leni Team, arXiv 2607.17044, 2026-07-19