Managed agent infrastructure becomes a product

Anthropic and LangChain released managed agent infrastructure within a day of each other. Developers now have to decide which parts of the harness they still want to build and operate themselves.

LangChain pairs deep agents with LangSmith

Harrison Chase on X

LangChain's new product combines its deep-agent system with managed LangSmith infrastructure. It arrived alongside Anthropic's managed-agent announcement, giving builders two concrete offerings to compare.

Build the harness or rent the runtime

Harrison Chase on X

The emerging stack separates agent logic from the authentication, sandboxing, and runtime services needed in production. Managed products package more of those services, while memory's place in the system remains unsettled.

Memory remains an open component

Harrison Chase on X

The managed-deep-agent discussion treats memory as a possible standard component rather than an application detail. The source doesn't establish whether vendors will bundle it into the runtime or leave it under developer control.

Claude Code sessions can message each other

Claude Code documentation

Claude Code now documents cross-session messaging, so separate coding sessions can coordinate directly. The capability creates a supported path for multi-session work instead of requiring an external coordinator.

Pi and Claude get an open-source link

Alon Wolenitz on X

A third-party extension connected Pi and Claude sessions on the same day the messaging documentation appeared. The quick follow-on shows how little work it takes to turn a session primitive into a multi-agent mesh.

A Codex auto-review loop consumed 6,000 turns

Faxan on X

An audit describes a Codex orchestration loop spending 6,000 turns in one day on repeated review. The report puts a concrete usage count on the cost of agent layers initiating work that an operator didn't request.

Rule Zero requires external verification

NakedPlato on Reddit

The proposed operating rule requires agents to verify their state outside their own reasoning before they continue. It gives teams a specific constraint they can add to session onboarding when autonomous loops are hard to observe.

DeepSeek's Terminal-Bench result gets a public-harness replication

Exciting-Camera3226 on Reddit

An independent public-harness run reports 82.7 percent on Terminal-Bench 2.1. It covered 445 trials, adding the trial count and replication that were absent from the vendor-reported result, although the unreleased minimal-mode harness remains a question.

A C99 BitNet engine reaches 36 tokens per second on a Xeon

shifu_legend on Reddit

A zero-dependency C inference engine reports 36 tokens per second for a 1.58-bit BitNet model on a Xeon processor. The implementation notes identify memory bandwidth as the ceiling for this local-inference setup.

SAP ties a travel and hiring freeze to AI spending

404 Media

SAP reportedly stopped most travel and hiring because of rising AI costs. The decision moves token spending from an infrastructure line item into staffing and operating policy at a major enterprise software company.

Miles Brundage points to a pending federal AI bill

Miles Brundage on X

Brundage argues that the current combination of incentives and law produces harmful outcomes, and he connects that concern to a specific federal proposal. His intervention follows the recent OpenAI and Anthropic incidents and comes from a former lab policy leader.

ChatGPT refuses direct requests for a named author's style

Ars Technica

ChatGPT now blocks direct requests to write in a named author's style while allowing requests for a similar feeling. The change sets a product policy boundary without a legal ruling requiring it.