Context Engineering with Deep Agents: How To Remember Without Keeping Everything In Context
Offloading, summarization, skills, subagents, and memory keep a long-running agent's context window focused instead of growing without bound.
AI Daily Recap
37 articles · 4 categories
The finishable daily brief
What happened in AI — Oct 14, 2026
Wednesday, Oct 14, 2026
37 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- LangChain: offloading, summarization, skills, subagents, and memory keep long-running agent context focused.
- Third Moment Research runs 10,000 agents for 36+ hours on GPT-6 Astra and GPT-6.1 Sol.
- vLLM adds a push-based NIXL connector that writes KV straight into decode memory over RDMA.
- BigQuery adds cost management for AI.GENERATE, AI.IF, AI.CLASSIFY, and AI.SCORE.
- Google's Antigravity SDK adds local model support.
- Meta extends Private Processing to AI glasses with confidential VMs and remote attestation.
Agent builders got practical patterns today: context engineering for long-running agents, a code review agent design, and 10,000-agent runs lasting 36+ hours.
Underneath, serving and spend controls are maturing, with vLLM's push-based KV transfer and BigQuery's cost management for AI functions, while Antigravity's local-model support pushes agents onto developer machines.
Agent context, code review, and reliability 5 items
Builders are publishing concrete patterns for keeping long-running and code-writing agents focused and checkable.
How to build a code review agent
Greptile walks through building a code review agent, a concrete reference design for the review step of an agent pipeline.
How Third Moment Research runs 10,000 agents for 36+ hours
Third Moment Research runs 10,000 agents for 36+ hours on GPT-6 Astra and GPT-6.1 Sol, a data point for multi-day agent runs that improve with use.
Amp puts AI agents to work beyond code with OpenAI models
Amp uses GPT-6 Astra, Ultrafast, and GPT-Live-1 to run agents beyond coding, with Sign in with ChatGPT for user auth.
Show HN: Agent-trials – a smoke-test runner for coding agent CLIs
A smoke-test runner for coding agent CLIs, aimed at catching regressions across agent versions.
Inference and AI cost control 4 items
Serving latency and spend are getting purpose-built controls, from KV transfer to SQL-level budgets.
Push-Based KV Transfer to Improve TTFT and TPOT for Disaggregated Inference in vLLM
vLLM's push-based NIXL connector has the prefill instance write KV directly into decode memory over RDMA, targeting TTFT and TPOT in disaggregated serving.
Predictable AI spend in BigQuery: Introducing cost management for generative AI functions
BigQuery adds cost management for AI.GENERATE, AI.IF, AI.CLASSIFY, and AI.SCORE so LLM calls inside SQL have predictable spend.
AMD MI355X joins AA-AgentPerf with DeepSeek V4.1 Flash results
AMD MI355X posts DeepSeek V4.1 Flash results on AA-AgentPerf, adding a non-Nvidia agent-workload data point.
Modal crosses $1B in Annualized Run Rate and raises $750M Series D
Modal crossed $1B annualized run rate and raised a $750M Series D led by Accel for AI application infrastructure.
Security, identity, and private execution 4 items
Security work this week centers on attested private compute, agent authorization, and faster malware triage.
Meta Extends Confidential Computing to AI Glasses with Private Processing
Meta extends Private Processing to AI glasses using confidential VMs, TEEs, remote attestation, anonymous credentials, and OHTTP routing.
SentinelOne cuts malware analysis from weeks to hours
With OpenAI Daybreak, SentinelOne cuts malware analysis from weeks to hours.
Show HN: AgentGate – Fast authorization gateway for AI agent tools
An authorization gateway that sits between AI agents and their tools.
Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection
DeepMind and GovTech Singapore fine-tune lightweight on-device Gemma models to catch conversational scams with no cloud data.
Local models and the open-weight gap 3 items
Agent SDKs are adding local execution while Chinese open models close in on the US frontier.
Google Antigravity SDK Adds Local Model Support for Greater Token Efficiency and Code Privacy
Google's Antigravity SDK now supports local models, so agentic workflows can run on a developer's machine for token efficiency and code privacy.
China's open AI models now trail the US frontier by just seven months
Reporting puts China's open AI models about seven months behind the US frontier.
Singapore office and government MOU | Anthropic
Anthropic opens its fifth Asia-Pacific office in Singapore and signs an MOU on AI safety, evaluation, and responsible adoption.
You are caught up for this edition