← All digests
AI Developer Digest

Thu, Nov 12, 2026

6 signals that cleared the gate18 min read
The Signal — start here
Thursday, November 12: the headliner is Anthropic completing its infrastructure tier model for Managed Agents — the platform now supports full self-hosting of the agent service itself (not just the tool-execution sandboxes), with your own PostgreSQL, your own inference provider, and your own identity layer. This is a major unlock for enterprise teams with data-residency or compliance requirements. The adjacent story is PydanticAI v2.68.0, which arrives with two platform-layer additions: GoogleDecisionsModel (connecting to Google's newly surfaced Gemini Decisions API) and SubAgents Temporal/DBOS child workflow support, extending the durable-execution surface further. Today's through-line is enterprise hardening — the agentic stack is maturing from "get it working" to "fit your compliance requirements."
Must-reads today
1
Anthropic Self-Hosted Managed Agents [HIGH]: the full service (API, agent loop, storage) now runs on your infrastructure; model calls go to any inference provider you choose — Bedrock, GCP, Microsoft Foundry, or your own gateway
2
PydanticAI v2.68.0 [MEDIUM]: ConversationStore for persistent UI conversations, GoogleDecisionsModel for the Gemini Decisions API, in-run user approval via RequestInput, and SubAgents child workflows for Temporal/DBOS

🏆 Repo of the Day

1

openai/codex-security — CLI + SDK that finds, validates, and patches security vulnerabilities in your codebase

npm install --global @openai/codex-security

cs login

Full repo scan

cs scan .

Scan only changed files against main

cs scan . --diff origin/main --mode deep


Breaking Changes

No breaking changes this period.


API & SDK Changes

1
High

Anthropic Self-Hosted Managed Agents — run the full Managed Agents service on your own infrastructure

What changed
Previously, Managed Agents had one self-hosting option — self-hosted sandboxes, where you run tool execution but Anthropic runs the agent service, session storage, and agent loop. Today Anthropic launched the other end of the spectrum: self-hosted Managed Agents, where you run the entire service — the API, agent loop, agents, sessions, and session history — on your own infrastructure, while model calls go to whichever inference provider you configure.
TL;DR
Anthropic's Managed Agents service (cma-server) is now deployable on your own infra (Linux/macOS, PostgreSQL 16+), with model calls routable to Bedrock, Google Cloud (Google-managed), Microsoft Foundry, or your own model gateway instead of Anthropic's API.
Developer signal
To request access to the cma-server download, contact your Anthropic account team. Once you have access: run cma-server alongside a PostgreSQL 16 database; create environments with the cma-server-admin CLI; run workers with ant beta:worker poll or the Python/TypeScript/Go SDK workers. Your applications and the ant CLI connect to your server's URL rather than api.anthropic.com — the API surface is identical, so existing code requires only a URL change. The inference credential (the API key / cloud credential your server uses for model calls) lives only in cma-server's config — it never reaches your applications or workers. Sandbox platform providers (Cloudflare, Daytona, E2B, Modal, Vercel, AWS Lambda, Fly.io, GKE) have published guides for running workers on their platforms. Important limitations: self-hosted Managed Agents is not eligible for ZDR (Zero Data Retention), does not support HIPAA, and some features available on the hosted service differ (see the Differences from the hosted service page). The agent's bash tool runs with the worker process's permissions — isolation is your responsibility; run workers in disposable containers or VMs.


Affects you ifYou build Managed Agents-based products and need agents, sessions, or session history stored on infrastructure you control; you need model calls routed to Amazon Bedrock, GCP (Google-managed), Microsoft Foundry, or your own gateway; your compliance requirements require data-residency guarantees that Anthropic's hosted service cannot provide.EffortSignificant (stand up cma-server, configure PostgreSQL, generate environment keys, wire an identity provider for the console, run workers in isolated sandboxes)

Research

arxiv.org DNS-blocked (eleventh consecutive day). huggingface.co/papers DNS-blocked (twelfth consecutive day). No research items confirmed from primary sources this period.


Tooling

1
Medium

PydanticAI v2.68.0 — ConversationStore, GoogleDecisionsModel, in-run RequestInput, Temporal cancellation

What changed
v2.67.0 → v2.68.0: adds ConversationStore for loading/saving UI adapter conversations across sessions; adds GoogleDecisionsModel backend for the Gemini Decisions API; adds RequestInput to pause a durable run and request approval or answers from a user mid-run; adds Temporal cancellation via a workflow signal; adds RefineOutput harness capability; enables SubAgents Temporal and DBOS child workflow delegation. Compatibility notes include renamed sub-agent event fields (adding Event suffix, with deprecated aliases kept) and harness parameter renames (num_results → max_results, max_duration_secs → max_duration_seconds, deprecated aliases kept). Also includes ~20 bug fixes covering model ID resolution, AnthropicModel.count_tokens sending betas, max_tokens clamping for Claude models, and compaction reporting.
TL;DR
PydanticAI v2.68.0 adds ConversationStore for persistent UI conversations, GoogleDecisionsModel for Google's new Gemini Decisions API, RequestInput for in-run human-in-the-loop approval, and Temporal workflow signal cancellation.
Developer signal
pip install --upgrade pydantic-ai to get v2.68.0. For ConversationStore: instantiate with ConversationStore(backend=...) and call store.load_conversation(session_id) to restore past UI adapter conversations into a new run — useful for resuming agent sessions with a history. For GoogleDecisionsModel: pass model='gemini-decisions-1' or instantiate GoogleDecisionsModel() directly — this requires a Gemini API key and targets Google's new Decisions API. For RequestInput: add RequestInput to your SubAgent task config to pause execution and prompt the user for a decision before proceeding; answered_by on the deferred call tracks who answered. The renamed event fields use deprecated aliases — existing code will continue to work but will log deprecation warnings; update num_results to max_results and max_duration_secs to max_duration_seconds before the aliases are removed in a future release. The ModelFallback behavior change (now moves to the next model when output hits max_tokens) may affect agents relying on partial truncated output from a fallback model — test your fallback chains.


Affects you ifYou use PydanticAI as your agent framework with Temporal or DBOS durable execution; you want to integrate Google's new Gemini Decisions API; you build multi-turn UI adapter conversations that need to persist across sessions; you rely on ModelFallback with any max_tokens-limited provider.EffortQuick for most features (drop-in upgrade); Moderate if you use renamed harness params or SubAgent event fields that now have deprecated aliases

Rising Dev Tools

1
Notable

openai/codex-security — CLI and SDK that automates the full vulnerability find-validate-patch workflow

Developer signal
Install with npm install --global @openai/codex-security, authenticate with cs login, then cs scan . --diff origin/main for a CI-safe PR scan. SDK usage: new CodexSecurity() + security.run("/path/to/repository"). Requires Node.js 22.13.0+. Some capabilities require Trusted Access for Cyber — verify your OpenAI account tier before adding to a production pipeline.


Affects you ifYou build or maintain applications and want to automate security review in CI; you're running LLM-generated code at scale and need automated vulnerability scanning with patches.
GitHub Trending (all languages, daily, Nov 12) | Stars: ~12.3k (↑ 663 today) | Link: github.com/openai/codex-security

Benchmarks & Leaderboards

lmarena.ai DNS-blocked (eleventh consecutive day). swebench.com DNS-blocked. No benchmark data confirmed from primary sources this period.


Technical Discussions

simonwillison.net DNS-blocked. hnrss.org inaccessible. Nothing cleared the quality bar this period.


Quick Hits

  • OpenHands v1.40.0 (Nov 12): improved Codex subscription authentication guidance, fixes for chat history at tied page boundaries, delayed-echo draft deletion, and git default-branch detection (stops assuming "main"). Source: github.com/OpenHands/OpenHands/releases/tag/v1.40.0
  • LangGraph 1.2.21 (Nov 12): fixes get_subgraphs to match whole namespace segments (not partial matches), and read_subgraph_namespace to read undeclared subgraph namespaces from the parent graph; security dependency bump (filelock + langgraph-api). Source: github.com/langchain-ai/langgraph/releases/tag/langgraph==1.2.21
  • langgraph-cli 0.4.43 (Nov 12): --deployment-aliases flag now accepts hidden aliases for Agent Server flags; removes the discontinued local debugger image from docker compose stacks. Source: github.com/langchain-ai/langgraph/releases/tag/langgraph-cli==0.4.43

Worth Watching (Announced, Not Yet Shipped)

  • Gemini API preview model deprecations — November 17 (5 days): Gemini 3.1 Flash TTS preview, Gemini Live API preview models, Gemini 2.5 TTS previews retire. Replacement: gemini-3.8-live and gemini-3.8-flash-tts. Source: ai.google.dev/gemini-api/docs/deprecations (DNS-blocked — verify at official URL).
  • Fast mode removal for Claude Opus 4.8 — November 30 (18 days): speed: "fast" with claude-opus-4-8 will error. Migrate to Claude Opus 5.5. Source: https://platform.claude.com/docs/en/release-notes/overview
  • Claude Sonnet 4.5 retirement — November 30 (18 days): claude-sonnet-4-5-20250929 retires. Migrate to Claude Sonnet 5.5. Source: https://platform.claude.com/docs/en/release-notes/overview


Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.