14 articles · 4 categories
The finishable daily brief
Monday, Sep 14, 2026
14 articles · 4 categories
read top to bottom · then stop
In 30 seconds
Agentic coding hit production limits on two fronts today: Anthropic disclosed its own CI system needed three rebuilds after agentic coding pushed job volume up 25x in six months, and a builder deliberately tested a coding agent against the classic prompt-injection "lethal trifecta" of private data, internet access, and a public repo.
Money kept moving through the open-weight side: Z.AI closed a roughly $5B share-and-bond raise for its next-generation GLM models, and Moonshot pushed out a Kimi K2.8 preview, extending a fast release cadence among Chinese labs.
Agentic coding's scale is now showing up as infrastructure strain, cost leakage, and rising autonomy — not just capability gains.
Anthropic's own CI job volume grew 25x in six months as agents wrote more code, and the team rebuilt its test-impact-selection service three times before it held.
A new tool audits agent transcripts for silent prompt-cache misses that spike Claude Code, Cursor, and Aider bills with no warning in the logs.
Perplexity now lets GPT-6 Astra write code, draft communications, and monitor production systems with far less human check-in than earlier models needed.
Z.AI's fundraising push (first reported as in-progress yesterday) closed, and open-weight labs kept shipping models on a fast, cost-driven cadence.
Z.AI's roughly $5B share-and-convertible-bond raise for next-generation GLM models has closed, as rising compute costs push Chinese labs toward public capital markets.
DeepSeek-V4.1-Flash packs 552B parameters behind an efficient MoE design, and DeepSeek has begun routing all V4-Pro API traffic to it at the cheaper Flash rate.
Moonshot rolled out a Kimi K2.8 preview across its Kimi Code and Kimi Work products, the latest in the rapid open-weight release cadence from Chinese labs this week.
Today's security threads span red-teaming agents against classic prompt-injection setups and building the audit trail and identity layer agents still lack.
Archestra deliberately gave a coding agent private data access, internet access, and a public repo at once — the classic prompt-injection "lethal trifecta" — to see what breaks.
METR and Redwood Research spent six days on-site at OpenAI producing an independent account of how its agents behaved during the earlier Hugging Face breach.
Kepil is a new open-source identity layer for AI agents — a passport, a mandate, and a tamper-evident action journal — aimed at making agent behavior auditable after the fact.
A new local catalog indexes roughly 3,000 community agent skills and runs a static risk scan on each before you install it.
Agents are moving deeper into regulated and revenue-critical workflows, from financial advising to healthcare to inbox management to GTM.
Anthropic launched Claude for Financial Advisors, connecting Claude to the custodians, portfolio platforms, and CRMs advisors use daily, plus skills built for that work.
Insight Health, Tennr, and Medallion described how they're building human-agent teams on Claude Tag inside their healthcare workflows.
Fyxer built its AI executive assistant on OpenAI models plus fine-tuning, memory, and real user feedback to organize inboxes and draft email in each user's own voice.
LangChain's own GTM agent lifted lead conversion 250% and saved each sales rep about 40 hours a month, per the team's write-up on building it.
You are caught up for this edition