The August brief
Change the contract before you change the stack.
Three capability families moved at once: frontier models expanded their context and tool surfaces, MCP moved its core toward stateless requests, and portable agent packaging arrived. The tempting response is a rebuild. I would resist it.
Keep the durable architecture from the AI Architect Guide 2026: explicit authority, application-owned state, bounded tools, task-specific evaluation, and evidence tied to each accepted outcome. Use this page as the dated delta layer.
Change ledger
Eight signals with an architectural consequence.
Dates identify the release or applicability point, not the day this page first noticed it. Follow each source before acting; vendor and regulatory surfaces can move.
OpenAI Assistants sunset; reusable Prompt objects next
OpenAI deprecations →Architecture consequenceAn application built on Assistants must move its state, tools, and generated items onto the current Responses stack, while reusable Prompt objects must move into application code before their shutdown.
Action before SeptemberInventory Assistant IDs and Threads. Map behavior to Responses, Conversations, and Items; export prompt content into reviewed, versioned application code rather than adopting a new Prompt-object dependency.
MCP became stateless at the protocol core
MCP 2026-07-28 release →Architecture consequenceConnection state can no longer carry workflow truth. Continuation, retries, authorization context, and durable state belong to the application.
Action before SeptemberTest protocol negotiation and re-issue behavior. Persist run state and idempotency outside the MCP connection.
The frontier model tier widened
Official model guides →Architecture consequenceLarge context, reasoning, tool use, and output limits now vary enough to justify task-level routing. A long context window still does not become durable memory.
Action before SeptemberRe-run workflow evals on one primary and one fallback route. Keep state, policy, and evidence outside the model.
Agent Plugins 1.0 standardized packaging
Agent Plugins specification →Architecture consequenceSkills and MCP servers can travel in one vendor-neutral package, while install, permission, policy, runtime, and user experience remain host responsibilities.
Action before SeptemberVersion portable procedures and tool servers. Keep authority policy in the receiving runtime, never inside the package claim.
EU AI Act Article 50 transparency duties entered application
European Commission guidance →Architecture consequenceFor in-scope EU use, disclosure and machine-readable marking can become product and content-pipeline requirements, not policy-page prose.
Action before SeptemberAsk counsel to scope applicability. Map any required disclosure, labelling, marking, and evidence to exact interfaces and generation paths.
Evaluation and telemetry interfaces are moving
Migration sources →Architecture consequenceProvider dashboards are replaceable surfaces. Your case set, grader logic, trace IDs, and release decisions must remain exportable.
Action before SeptemberMove new OpenAI evaluation work to current datasets, traces, and agent-eval surfaces; move legacy Langfuse trace ingestion to OTLP/HTTP before its cloud sunset.
Zero Data Retention expanded for eligible OpenAI API customers
OpenAI data controls →Architecture consequenceRetention is now a feature-by-feature architecture property. Background work may still need temporary response persistence for polling.
Action before SeptemberDraw the actual data path for every API feature. Record eligibility, storage behavior, region, retention, and deletion evidence in the vendor decision.
Official eval-sandbox incident reviews exposed infrastructure risk
Official incident reviews →Architecture consequenceAn evaluation harness can touch credentials, proxies, and networks. A prompt-level boundary cannot contain a misconfigured sandbox.
Action before SeptemberEnforce egress, credential, identity, filesystem, and audit boundaries in infrastructure. Treat evaluation environments as production security surfaces.
Band 01 · Act now
Close the migrations and authority gaps.
These items already have a sunset, an applicability date, or a disclosed incident.
Remove Assistants dependencies.
OpenAI’s migration guide states that the Assistants API is no longer available as of 26 August 2026. The durable replacement separates Responses, Conversations, and generalized Items. Do not make reusable Prompt objects the new dependency: OpenAI’s deprecations page schedules the v1/prompts API and reusable Prompt objects to shut down on 30 November 2026 and directs teams to move their content into application code.
Preserve behavior before changing APIs: record the current tools, instructions, thread state, files, output handling, and failure paths. Then migrate one eval-backed workflow slice at a time.
Read the official migration guide →Turn transparency duties into product controls.
The European Commission’s Article 50 guidance says the relevant transparency obligations apply from 2 August 2026. Applicability depends on the system, role, content, and jurisdiction.
For in-scope work, disclosure and machine-readable marking need owners, interfaces, content-pipeline controls, and evidence. This field note is architecture guidance, not legal advice; scope the duty with qualified counsel.
Review the Commission guidance →Harden the evaluation sandbox.
Anthropic and OpenAI published separate incident reviews around cybersecurity evaluation infrastructure in this release window. The practical lesson is narrow: evaluation environments can become real attack surfaces when network, credentials, identity, or proxy boundaries are wrong.
Put containment in infrastructure: deny-by-default egress, short-lived scoped credentials, isolated filesystems, explicit proxy policy, audit logs, and a tested kill path.
Own the eval set and trace schema.
OpenAI’s legacy Evals platform becomes read-only on 31 October and shuts down on 30 November 2026. Langfuse says its legacy cloud trace-ingestion API sunsets on 16 November 2026 in favor of OTLP/HTTP.
Export cases, grader definitions, historical results, trace identifiers, and release decisions now. A dashboard can change; the evidence behind a production decision must remain inspectable.
Band 02 · Adopt and test
Use the new capability without handing it the architecture.
These releases are usable now. Promotion still belongs to your workflow evaluation, privacy boundary, and operating envelope.
Test reasoning continuity and task-tier routing.
The current model catalog documents GPT-5.6 Sol, Terra, and Luna with 1.05 million tokens of context and up to 128,000 output tokens, plus current guidance for programmatic tool use and persisted reasoning.
Official guide →Test long-horizon work with explicit compaction.
Anthropic documents a one-million-token context window, up to 128,000 output tokens, thinking enabled by default, and beta mid-conversation tool changes. Keep durable state outside that context.
Official platform notes →Test the workhorse route on your own cases.
Google lists Gemini 3.7 Flash as generally available from 13 August 2026, with a one-million-token context window, up to 64,000 output tokens, tunable thinking, and built-in tool support.
Official guide →MCP 2026-07-28
Make state ownership visible.
The current MCP revision removes the initialization handshake and makes each request carry its protocol version and client capabilities. That improves horizontal reasoning about the protocol boundary while moving continuation truth into explicit application data.
- Persist workflow state independently of a connection.
- Use idempotency around every retried side effect.
- Test capabilities and version behavior on each request path.
- Keep authorization and consent decisions outside model output.
Agent Plugins 1.0
Package procedure and capability, not permission.
Agent Plugins 1.0 provides a vendor-neutral package for Agent Skills and MCP servers. The specification deliberately leaves installation, permissions, policy, runtime behavior, and UX to the host.
- Version Skills as operational procedure.
- Expose narrow, typed MCP capabilities.
- Require host-side permission and policy review.
- Verify the exact package before admission.
Data controls
Evaluate retention per feature.
OpenAI announced Zero Data Retention availability for eligible frontier-model API customers on 19 August. Its data-control documentation also notes that background mode needs temporary response persistence for polling.
The architecture task is a feature-level data-flow record: what leaves your boundary, where it is processed, what persists, which controls apply, and what evidence proves deletion or expiry.
Check current data-control behavior →Band 03 · Watch
Track the standards that are still moving.
Watch items belong in a decision log, not in a production dependency by default.
OpenTelemetry GenAI conventions
OpenTelemetry is the portability direction for model, agent, tool, and MCP telemetry. The GenAI conventions continue to evolve. Pin the schema version emitted by your instrumentation and preserve your workflow-specific evidence fields.
Follow the official repository →Anthropic Model Hardware Standard preview
Anthropic introduced a research preview on 27 August for a model-agnostic driver layer connecting agents to programmable scientific and manufacturing devices. It is a research signal, not an open production standard.
Read the research-preview announcement →Regional and open-model topology
Regional inference and open-model options are expanding. Track data locality, inference jurisdiction, model rights, operational ownership, and exit portability as separate decisions. A single umbrella label hides those trade-offs.
Review Mistral’s August announcement →What did not change
The durable architecture still owns the outcome.
Human authority remains explicit. A larger model does not inherit the right to approve spend, publish, merge, message, or change a system of record.
Tools remain typed and narrow. A portable package or protocol does not make a capability safe, authorized, or appropriate for the current workflow.
State remains application-owned. Context windows, sessions, and connection metadata cannot replace durable business records and replayable events.
Evaluation remains workflow-specific. Vendor benchmarks do not test your recipient, data, tools, policy, or failure modes.
One recurring workflow is still the right unit. It gives the team a bounded recipient, accepted result, authority map, and evidence set.
August architecture audit
Twelve questions for the next review.
Run this against one workflow. Each “no” becomes an owned task, a recorded exception, or a reason not to migrate yet.
- 01
Can we replace the primary model route without changing business workflow code?
- 02
Does every long-running job persist state outside model context and transport connections?
- 03
Have we tested the MCP version, capability, retry, and idempotency paths we actually deploy?
- 04
Do tool packages declare capabilities without granting themselves authority?
- 05
Can an operator trace one accepted outcome through model, tool, policy, approval, and effect events?
- 06
Are our eval cases and grader definitions exportable from the current vendor surface?
- 07
Have retention and residency been checked for each feature, including background execution?
- 08
Are evaluation sandboxes isolated by infrastructure controls with no ambient credentials?
- 09
Does the interface disclose AI interaction or generated content where our scoped obligations require it?
- 10
Can a human pause, reject, repair, and roll back a consequential run?
- 11
Did we re-run representative, adversarial, timeout, and stop cases after the August changes?
- 12
Which migration or watch item has a named owner and a date for the next decision?
Source ledger
Primary sources checked 28 August 2026.
Product and policy details are snapshots. Recheck the linked authority before a migration, purchase, compliance, or production decision.
Models and runtime
Protocols and packaging
Migration, data, and governance
Method: this edition uses official documentation, specifications, provider announcements, incident reports, and regulator guidance. It does not convert vendor benchmarks into workflow claims. Legal applicability remains a scoped counsel decision.
Connected guide
Use the evergreen system, then apply the August delta.
The full guide defines the seven architecture planes, decision records, evidence gates, operating artifacts, and 30-day workflow plan. This edition only tells you what moved.