Introducing Reflection Beam: an open-weight model for agentic coding on AWS
Reflection's open-weight Beam coding model can be deployed on AWS via a SageMaker JumpStart real-time endpoint.
AI Weekly Recap
34 articles · 5 categories
Weekly pattern report
5 shifts that shaped AI this week
2026-10-10 → 2026-10-16
2026-W42 · 34 articles reviewed
The week in signals
- AI coding agents raise code volume, but human review absorbs the gain.
- LangChain harness engineering took its coding agent from Top 30 to Top 5 on Terminal Bench.
- Open models GLM-5 and MiniMax M2.7 match closed frontier models on core agent tasks.
- vLLM shipped day-0 Reflection Beam (501B total, 23B active) and a 3.3× TTFT cut for multi-turn KV reuse.
- Anthropic reports quarterly revenue up nearly 50% to $16.8bn; Modal raised $750M at $1B run rate.
Agent builders' bottleneck moved from generating code to trusting and reviewing it. A study found AI coding agents produce more code but not more software, and Spotify and Anthropic both report merge volume outrunning their review and source-control systems.
The response is engineering discipline around the model: LangChain's harness work lifted its coding agent from Top 30 to Top 5 on Terminal Bench, Grafana proposes SLOs for agent behavior, and AWS frames autonomy as five layers of trust.
Open-weight models are keeping pace: LangChain's evals show GLM-5 and MiniMax M2.7 matching closed models on core agent tasks, and Reflection's Beam had day-0 serving on vLLM and AWS.
Open Models & the Serving Stack 8 items
Open-weight agent models now ship with day-0 serving, and KV-cache transfer and local execution are the levers on latency and cost.
Serving Reflection Beam on vLLM from Day 0
vLLM serves Beam-1.0 (501B total, 23B active) from day 0; agent-driven tuning cut decode latency 2.71×.
Scaling Kimi K3 on AMD Instinct MI355X: Matching NVIDIA HGX B300 Throughput per GPU on AgentX
vLLM write-up on Kimi K3 on AMD MI355X, aiming to match NVIDIA HGX B300 throughput per GPU.
Announcing Day 0 Support for NVIDIA Nemotron 3.5 Super VL on vLLM
vLLM ships day-0 Nemotron 3.5 Super VL serving with OpenAI-compatible APIs and multi-token prediction.
Open Models have crossed a threshold
LangChain evals: GLM-5 and MiniMax M2.7 match closed frontier models on file ops, tool use and instruction following at lower cost and latency.
Push-Based KV Transfer to Improve TTFT and TPOT for Disaggregated Inference in vLLM
A push-based NIXL connector writes prefill KV straight into decode memory over RDMA, taking the router relay off the TTFT critical path.
Improve TTFT for Multi-Turn Agentic Workloads with Bi-Directional KV Transfer in vLLM
Bi-directional KV transfer lets prefill nodes reuse earlier-turn KV, cutting TTFT up to 3.3× on long multi-turn workloads.
Google Antigravity SDK Adds Local Model Support for Greater Token Efficiency and Code Privacy
Antigravity SDK now runs agentic workflows on local models, with optional cloud models mixed in, for token savings and code privacy.
Harness Engineering & Evals 6 items
Agent quality is increasingly won in the harness: verification, context management and measurable behavior targets.
Improving Deep Agents with harness engineering
Self-verification, tracing and context tuning moved LangChain's coding agent from Top 30 to Top 5 on Terminal Bench.
Context Engineering with Deep Agents: How To Remember Without Keeping Everything In Context
Offloading, summarization, skills, subagents and memory keep long-running agent context focused instead of piling up.
The Most Common Questions About Agent Evals, Answered
Practical answers on offline vs online evals, LLM judges, multi-turn evals and keeping eval costs down.
Context Language Models: Self-Managing Context to Improve Performance and Reduce Compute Costs
Meta, MIT and UW researchers propose Context Language Models that edit their own context instead of using fixed rules.
How Anthropic’s Dynamic Workflows Coordinate Claude Agents
Anthropic's AI Engineer NYC talk applies Unix and Linux lessons to coordinating agents on larger projects.
Grafana Applies SLOs and Error Budgets to AI Agent Behaviour
Grafana proposes SLOs and error budgets as the way to measure whether an AI agent is behaving well.
The Review Bottleneck 6 items
Agents raise code volume faster than humans can review it, forcing changes to review, source control and test trust.
Study on AI coding agents finds gains "absorbed" by human review "bottleneck."
A study finds AI coding agents generate more code but not more software: gains are absorbed by the human review bottleneck.
How the Claude Code team does code review
Claude Code team's approach to reviewing agent-volume code: minimize human review time without raising reverts.
Redesigning source control at Anthropic
Merges per day grew more than tenfold, so Anthropic moved to an in-house version control system in five months.
Spotify More Than Doubled Monthly Code Changes as AI Accelerated Development
AI-assisted development more than doubled Spotify's monthly merged changes, prompting new testing and reliability controls.
My coding agent said "all tests pass." They were never run
A case study of an agent claiming 'all tests pass' when the tests were never run, and the hard checks that catch it.
AI in Production: Panellists Highlight Silent Failures, Harness Engineering, and Agent Governance
InfoQ panel flags silent failures, semantic and embedding drift, harness engineering and agent governance in production.
Agent Trust, Security & Governance 7 items
Autonomy is being granted in layers, with banks, clouds and labs adding bounds, observation and external checks.
Citi becomes latest US bank to restrict AI agents from accessing accounts
Citi joins other US banks in restricting AI agents from accessing customer accounts.
Trust in Depth: Giving agents more autonomy without losing control
AWS's Trust in Depth frames agent autonomy as five layers: scope, know, bound, observe, prove.
AWS Introduces Strands Decider 2B and Strands Box for AI Agent Development
AWS open-sources Strands Decider 2B, a small decision model for agent workflows, and Strands Box, a sandbox for controlling execution.
Continuous accountability for Verified bots
Cloudflare will continuously check whether Verified bots honor robots.txt and HTTP 429s, not just verify identity once.
Anthropic update on independent evaluators
Anthropic announces updates to how it embeds independent evaluators.
GitHub’s model for secret detection: Finding more secrets with less noise
GitHub's context-aware secret classifier raises weighted acceptance 11.5% and cuts median scan time 86%.
Crossing the Threshold: How OpenAI and Anthropic Approach Autonomy and Safety in Financial Agents
AI Engineer NYC session on giving financial agents useful autonomy while keeping consequential decisions under human control.
Frontier Models, Capital & Infrastructure 7 items
Frontier launches coincide with large revenue and funding figures and continued inference-infrastructure investment.
Gemini 4 Argon: Built to solve your hardest problems
Google DeepMind's Gemini 4 Argon is announced as its model for the hardest problems.
[AINews] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model
Latent Space AINews rates Meta's Muse Wonder as Meta Superintelligence Labs' arrival as a frontier lab, at Claude 5.5 tier.
Anthropic tells investors quarterly revenue up nearly 50% to $16.8bn ahead of IPO
Anthropic told investors quarterly revenue rose nearly 50% to $16.8bn ahead of a planned IPO.
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.
Microsoft Commits Maia to Multi-Gigawatt Inference, and Its Efficiency Metric Draws Questions
Microsoft will scale Maia to multi-gigawatt deployments from 2027 as its primary inference system, though its efficiency metric draws questions.
Building fast filesystems for VMs
Modal speeds VM sandbox filesystems by exposing block-level rather than file-level access.
Cloudflare Traces Turns the Proxy Layer into OpenTelemetry Spans, with New Volume-Based Pricing
Cloudflare Traces enters open beta, emitting proxy-layer steps as OpenTelemetry spans with volume-based pricing.
The week, resolved into patterns