Weekstart Brief — Issue 1

1. OpenAI Agents API — sandbox billing + ZDR/US limits

No separate Agents platform fee. Hosted sandboxes bill at container rates (1/4/16/64 GB → $0.03/$0.12/$0.48/$1.92 per 20-min session unit), eligible sessions per minute with a 5-min minimum. Model + tool fees extra. Agents API is US data residency only and does not support ZDR — self-hosted does not make it ZDR-eligible.

Why it matters: “No Agents fee” is not free compute, and compliance assumptions from Chat/Completions won’t carry over.

Action: Pin container tier, reuse long-lived containers, gate Agents API behind US-residency / non-ZDR policy before prod. Wait for environment connected before file ops (1h idle expiry on hosted).

Links: Agents overview · Hosted environments · Pricing

2. Assistants API is still dead — finish the cutover

Sunset 2026-08-26. Assistants→Prompts, Threads→Conversations, Runs→Responses. No automated Threads→Conversations migration.

Why it matters: Lagging SDK paths fail hard Monday morning.

Action: Cut over to Responses + Conversations (or Agents API for durable work). Rebuild history from app-stored messages. Prefer prompt specs in source control — dashboard prompt objects are on a deprecation path too.

Links: Migration guide · Deprecations

3. Gemini Antigravity default → gemini-3.8-flash

Managed-agent default moved to Gemini 3.8 Flash. Saved agents lock model at agents.create (interaction-time override blocked).

Why it matters: Unpinned runs silently changed cost and behavior.

Action: Set agent_config.model explicitly on every create; audit existing agents; set max_total_tokens to stop runaway loops.

Links: Antigravity agent docs · Managed agents expansion

4. Cursor Projects — no separate quota

Projects burn Cloud Agent / Cursor Models / Other Models pools. Subscriptions and automations are always-on meters.

Why it matters: Parallel subagents can exhaust the month and trip on-demand charges with no “Projects limit” line item.

Action: Set spend limits before Monday rollouts; watch the Spending tab; prefer Cursor Models pool where possible.

Links: Projects · Models & pricing · Usage limits

5. LangChain Deep Agents v0.7 — lean harness, breaking defaults

~65% fewer base input tokens; todos opt-in; backend factories removed; ls/glob/read_file output shapes changed; delete added to default FS tools.

Why it matters: Teams on ≤0.6 will see missing todos and broken parsers after upgrade.

Action: Upgrade with a migration checklist — restore TodoListMiddleware if needed; pass concrete backends + explicit StoreBackend(namespace=...); re-parse FS tool outputs; allowlist FS tools if you don’t want delete.

Links: Deep Agents v0.7 · Changelog · Release

Agents API environment shape

Choose environment.type: openai_hosted vs self_hosted; network enabled|disabled|restricted; packages / setup_commands / files. Header: OpenAI-Beta: agents=v1.

Anthropic MCP tunnels

Beta mcp-tunnels-2026-06-22; WIF scope workspace:manage_tunnels. IDs: tnl_*, tcrt_*. Max 2 active CAs. Admin API keys rejected. Pair Messages/Managed Agents with tunneled URL + mcp-client-2025-11-20.

Gemini Antigravity agent_config

{ type: "antigravity", model, max_total_tokens }. MCP tools: type: "mcp_server", lowercase alphanumeric name, Streamable HTTP only. Hooks: matcher regex + deny JSON {"decision":"deny","reason":"..."}.

Deep Agents 0.7 filesystem / store

Removed BackendFactory / BACKEND_TYPES / FileFormat / Unset. FileData.content is a string. Empty ls/glob → "No files found". No cat -n gutter on reads. Pass explicit StoreBackend(namespace=...).

Claude Fable/Mythos 5.1 tools

Prefer strict tool use or structured outputs (output_config.format). tool_choice only auto/none. Thinking blocks are model-bound on replay.

Google UAP ↔ MCP

CEL on destination.agent_registry.mcp_server.name, .tool.name, .tool.annotations.destructive_hint / read_only_hint. Principals as SPIFFE agent identities. Deny overrides allow. No VPC-SC for this feature.