Golden Data

Issue #15 · August 16, 2026

They deprecated the part that worked.

Paul BrownWrites Golden Data from Ogden, Utah6 min read

Every Sunday I look for the sentence a vendor writes when it wants a change to sound like nothing happened. AWS published a good one this month, in the Bedrock documentation, in the FAQ entry about the new name: “This is a naming update only. No action is required on your part.”

Then read the rest of the page. Amazon Bedrock Agents, launched November 2023, became Amazon Bedrock Agents Classic on July 30 and closed to new customers the same day. The model catalog froze on that date. No new features are planned. An AWS account with no Bedrock Agents activity in the previous twelve months now receives an HTTP 403 on CreateAgent, and the FAQ is direct that there is no exception process.

Nothing broke. Existing agents run, every API still answers, the bill is unchanged. What ended is the road forward for one specific layer — the orchestration loop. Pre-processing, tool selection, return of control, the trace. Not the model, which keeps getting better on the other side of the wall. The middle.

Ten days after this issue lands, OpenAI removes the Assistants API on the same reasoning. Vendors are retiring the middle. A paper posted July 19 puts a number on what the middle is worth, and the number is most of it.

What’s actually shipping this week

AWS, July 30. Bedrock Agents Classic is in maintenance mode with no announced end-of-life and no migration deadline. Accounts that used it in the past year keep it; everyone else gets the 403. The successor, AgentCore, is a different shape: a config-based harness, or any framework you like — Strands, LangChain, the OpenAI Agents SDK, the Claude Agent SDK — deployed on managed runtime. The migration comparison table is where the honesty is. Stage-specific prompt overrides are “not directly replicated.” Routing-mode multi-agent is “not straightforward today.” The recommended migration tool is itself an agent skill you install and then prompt: “Help me migrate my Bedrock Agent to AgentCore harness.”

The pricing shape underneath it. There is no charge for Bedrock Agents Classic itself — you pay for inference and whatever Lambda and Knowledge Base resources it calls. AgentCore is consumption-priced across runtime, memory, and gateway, with no separate charge for harness orchestration, and AWS notes its prompts are more token-efficient, so inference may come out flat or lower. Total cost could genuinely fall. The billable surface still went from one line to four.

OpenAI and Microsoft, August 26. OpenAI notified Assistants API users on August 26, 2025 and removes it exactly twelve months later, with the Responses API as the destination. Azure inherits the same date for the Azure OpenAI Assistants API, and Microsoft Foundry Agent Service (classic) carries its own retirement on March 31, 2027. Twelve months is now roughly the notice period for agent infrastructure. Enterprise software is still bought on three-year paper.

Microsoft Copilot Studio, as documented on June 15. For agents created before March 18, 2026, moving to Microsoft Entra Agent ID is not a migration. The docs say it plainly: “there is no automated or in-place migration path to convert an existing Copilot Studio agent’s service principal to an agent identity.” You build a new agent, rebuild topics, knowledge sources, permissions, connectors and channel deployments by hand, run both for 10–14 days, then decommission the old one. For high-usage production agents the guidance is to wait — don’t recreate them until Microsoft ships an official automated path. Google ran the same clock on a smaller surface: the generative modules in the Vertex AI SDK were deprecated June 24, 2025 and removed June 24, 2026.

The research, July 19. “Where Does Agent Reliability Come From? A Cross-Benchmark Decomposition of Verification Loops, Specialist Models, and Scaffolding in a Production Enterprise Agent” (Arunabh Dastidar, for the Leni Team; arXiv 2607.17044) decomposes a production business-analyst agent against its frontier base model: 91.25% versus 80.25% on SpreadsheetBench Verified (n=400, p<0.001), and 75.2% pass@1 on the GAIA validation split (n=165). The attribution is the finding. Most of the gain comes from scaffolding, routing and specialist models; the verification step in isolation contributes +1.5 points. This is one vendor measuring its own system, and it is not the systems above. The mechanism is shared, and it is the one being deprecated: the accuracy lives in the loop around the model, not in the model.

What I’d ship in your app this week

Build the platform-dependency ledger. A vendor’s maintenance-mode notice should not be the first time anyone can answer how many agents you have on that platform and what breaks if it moves.

  • User moment. The email arrives at 9am. By 9:30 someone has to tell an exec how exposed you are, and today that answer takes a week.

  • Data shape. One row per agent per dependency: (agent_id, owner, env, platform, surface, surface_kind, portable_equivalent, rebuild_days_est, last_invoked_at, invocations_30d, on_revenue_path). surface_kind is the column that earns its keep — orchestration, memory, identity, tool transport, eval.

  • System shape. A nightly job, not an agent. Read the platforms’ own list APIs, join to your IaC, and enrich from the fingerprints the vendors leave: ListAgents on Bedrock, the AgentCreatedBy:CopilotStudio service-principal tag in Entra.

  • Cost and latency. Engineer-days, not tokens. Two days to build, minutes to run, no request path touched.

  • Eval shape. The ledger is correct when it predicts which of your agents a vendor notice touches before the vendor’s console tells you. Test it against the July 30 notice retroactively.

  • Instrumentation. Alert on agents created outside the ledger. That number is your real drift rate.

  • Two weeks in. Success is a one-screen answer with a rebuild estimate attached. The credible failure is a stale inventory nobody believes, which is why it derives from live APIs and never from a spreadsheet.

Then the smaller piece, and do this one first if you only do one: freeze a hundred-task eval set from your own production traffic, record today’s pass rate, and keep it. When you rebuild scaffolding, the paper above says you are rebuilding the layer your accuracy came from. “It still answers” is not a measurement.

Give the buyers their due. Every one of these successors is more open than the thing it replaced. AgentCore runs frameworks AWS doesn’t own and models AWS doesn’t sell. The Responses API brought MCP with it. Entra Agent ID puts agents under Conditional Access and a real audit trail instead of a service principal wearing a costume. You lost a managed loop that was free, and you gained a loop you own well enough to take somewhere else. Ask what your vendor’s twelve months would cost you, before the twelve months start.

Sources

Send me an email and we will talk. If something here landed close to what you're working on, the door is open. No calendar funnel, no pitch deck — I read every note that comes in.

Doing the work rather than deciding what to build? Crafting is the column for that chair.

Email meRead more Signal

← All Signal issues · Drafted with Claude · Edited by Paul Brown