Agentic Business

FINANCE STACK

Anthropic stops selling shovels, starts selling mines

via Anthropic

For two years, frontier labs sold APIs and let everyone else figure out the workflow. Today Anthropic just shipped the workflow.

The lab dropped 10 ready-to-run finance agent templates covering pitchbook building, KYC screening, and month-end close, available as plugins inside Claude Cowork and Claude Code, and as cookbooks for Claude Managed Agents. They paired it with a 64.37% score for Claude Opus 4.7 on Vals AI's Finance Agent benchmark.

The connector story is the louder shoe drop. Eight new data integrations landed alongside the templates, plus a Moody's MCP app that pulls from a 600M+ company dataset.

• Templates ship inside Claude Cowork, Claude Code, and Managed Agents cookbooks
• New connectors: Dun and Bradstreet, Verisk, Guidepoint, IBISWorld, plus Moody's via MCP
• Claude add-ins now span Excel, PowerPoint, Word, and Outlook

Why it matters: Vertical agent recipes from a frontier lab mean finance and insurance teams can clone working agents in days instead of stitching connectors together for a quarter. Read more.

MODEL SHUFFLE

GPT-5.5 Instant quietly slides into the driver's seat

via OpenAI

OpenAI didn't throw a launch event. It just swapped the engine while you were still driving.

GPT-5.5 Instant is now the default model in ChatGPT, and the headline number is a 52.5% drop in hallucinated claims compared to GPT-5.3 Instant on high-stakes prompts in medicine, law, and finance. It's also live in the API as chat-latest, with GPT-5.3 Instant scheduled to retire in three months.

If you've been pinning chat-latest in production, congratulations, you've been migrated. The accuracy claim is squarely aimed at the regulated-domain agents that teams have been holding back from real deployments.

• New default in ChatGPT for free and paid tiers
• Available in API as chat-latest
• GPT-5.3 Instant retires in three months

Why it matters: The "we can't ship agents into compliance-heavy domains because they hallucinate" excuse just got a fresh rebuttal you have to test against. Read more.

RAPID FIRE

OpenAI shipped its Agents SDK twice before lunch

via GitHub

If your dependabot got chatty this morning, that's why. OpenAI cut two Agents SDK releases in roughly 12 hours.

v0.15.2 introduced a first-class "context management model setting," giving long-running agents an explicit knob for context strategy instead of the duct-tape approach most teams have been bolting onto their loops.

v0.15.3 followed with MCP hardening: tool input schemas can no longer be mutated, non-object JSON gets rejected, duplicate tool registrations now error deterministically, and an audio-format-negotiation race condition got patched.

What this fixes in production

• That weird MCP bug where the same agent answered differently across pods? Probably the duplicate-tool race.
• Long-running agents that ballooned past the context window now have a sanctioned strategy hook.
• Audio agents that occasionally negotiated the wrong codec are unstuck.

Why it matters: The MCP fixes remove a class of nondeterministic tool-call bugs that have been quietly biting production deployments all year. Read more.

IDENTITY UNLOCK

Anthropic's SDK finally speaks enterprise SSO

via GitHub

The single biggest reason regulated orgs couldn't put Claude agents into prod was static API keys. That blocker just got smaller.

anthropic-sdk-python v0.99.0 adds workspace targeting for OIDC federation token exchange. It builds on v0.98.0 from May 4, which introduced Workload Identity Federation, interactive OAuth, and auth profiles, plus updates to the Managed Agents API.

Translation for your CISO: agents can now authenticate via OIDC-issued tokens scoped to a specific workspace. No long-lived secrets sitting in env vars, no service accounts being shared across teams.

• Workspace-scoped OIDC token exchange in v0.99.0
• Workload Identity Federation and interactive OAuth shipped in v0.98.0
• Refreshed Managed Agents API surface

Why it matters: Per-workspace federated identity through the official SDK clears one of the last enterprise-procurement hurdles for Claude-based agents. Read more.

PATCH TUESDAY

LangChain plugs the hole the security crowd has been circling

via GitHub

If you persist agents to disk and rehydrate them later, stop reading and go upgrade. We'll wait.

langchain-core 1.3.3 and langchain 0.3.29 landed as security releases. The load() path is hardened against untrusted manifests, and langchain.storage._lc_store now restricts deserialization.

This closes a remote-code-execution surface that attack research has been circling all year. Teams that serialize chains or agents into Redis, S3, or Postgres and rehydrate them at runtime are the exact target shape.

• langchain-core 1.3.3 hardens the core load path
• langchain 0.3.29 restricts deserialization in the storage backend
• Pin bumps belong in your next deploy, not next sprint

Why it matters: If your agent infra rehydrates serialized state from any persistence layer, this is the upgrade you don't get to defer. Read more.