Redesigning source control at Anthropic \ Anthropic
## Anthropic describes Scion, an in-house source-control system for agent-scale development
source-linked1
Signals
What several Field Notes show together. A signal is written only when two or more notes converge on one meaning, in a few sentences with links to the notes behind it.
Latest signals
Archive: raw signals
Single dated, source-linked observations recorded before 2026-10-21, kept as evidence records rather than conclusions.
Key takeaway
Across 1742 observations, most signals relate to model behavior, evaluation discipline, and real-world agent operations. Strong evidence clusters around tooling, memory, routing, and productized agent execution.
agentscoding agentsagent harnessagent runtimemcpmodels
What this means
Teams are moving from experiments to systems. The emphasis is on verifiable behavior, operational safety, and measurable outcomes—especially in agent memory, evaluation rigor, and forward-deployed engineering.
Filters (current page)
1-10 of 10 matching raw signals on this page
thropic-dcb9ce notableresearch
source-linked1-models-20e9ea notableresearchRethinking Confidence in Reasoning Models
Cohere describes CALIBER, a training method that asks a reasoning model to report two confidence estimates: its expected chance of solving a problem before reasoning and its estimated chance that the particular answer is correct afterward.
source-linked1asoning-b32f20 notablereleaseShared context between GPT-Rosalind 's scientific reasoning and program-level decisions.
OpenAI says GPT-Rosalind, its specialized life-sciences model, is now globally available to eligible organizations through its trusted-access program, moving out of research preview.
source-linked1thropic-7ba71b notableannouncementStrengthening our defenses against distillation attacks \ Anthropic
Anthropic said it has expanded defenses against attempts to extract Claude’s reasoning for model distillation, combining controls on reasoning context, a two-stage classifier, and stricter limits and bans for persistent attempts.
source-linked1-s-open-0983dc notablereleaseUnifying Edge AI: How Coral NPU’s open standard brings portable, energy-viable intelligence from model to silicon
Google Research has completed an open 32-bit Coral NPU architecture based on RISC-V, adding a Matrix Execution Unit.
source-linked1tyv-bio-9efa7e notableresearch211 designs were lab-tested at Adaptyv Bio and Twist Bioscience ; 17 met strict binder criteria.
Anthropic has released a public dataset of 211 lab-tested protein designs targeting scorpion toxins.
source-linked1s-the-2-745e0f notableresearchALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model
IBM Research describes ALTK-Evolve, a method that distills behavioral guidelines from an agent’s training-task attempts and appends them to its system prompt without changing model weights.
source-linked1gations-2a1994 notablereleaseFactory now supports @Linear delegations.
Factory’s Linear delegation documentation describes an issue-level entry point for Droid.
source-linked1-issues-098194 notableupdateHow to review your agents for security issues (by Anthropic): Learn how to audit finished agent sessions for things like prompt injection, unauthorized hosts, leaked credentials, blocked calls that...
Anthropic published a Claude Cookbook workflow for reviewing completed agent sessions for security issues.
source-linked1backend-fe9f6f notablereleaseHugging Face announced TorchTPU, a native PyTorch backend that adds Google TPU as a device and extends TPU
Hugging Face announced TorchTPU, a native PyTorch backend that adds Google TPU as a device and extends TPU support across Transformers, Diffusers, and Kernels.
source-linked1
No raw signals match these filters.
Source & exports
Primary indexes and programmatic access to this dataset.
Archive contractEnglish normalized copy, Telegram provenance, public sources only