We built one harness for every agent, inspired by Ramp's Inspect
Metaview runs sourcing, screening and reporting agents on one shared harness, modeled on Ramp's Inspect, instead of per-agent scaffolding.
AI Daily Recap
24 articles · 5 categories
The finishable daily brief
What happened in AI — Nov 3, 2026
Tuesday, Nov 3, 2026
24 articles · 5 categories
read top to bottom · then stop
In 30 seconds
- Claude Platform ships native OpenTelemetry ingestion into Databricks MLflow.
- LangSmith BYOC is GA on AWS, Beta on Azure and Google Cloud.
- Agent safety evals may not predict deployment behavior because of eval-awareness.
- Mandiant: stolen AI API keys produced nearly $1M in charges for one company.
- 'Weight Space Programming' paper: LLMs can read and write network weights.
- MoE inference for Apple Silicon fits 125B parameters in under 10 GB of peak memory.
Observability went OpenTelemetry-native: Claude Platform traces now flow into Databricks MLflow, and LangSmith BYOC reached GA on AWS with Azure and Google Cloud in Beta.
Alongside it, a study of eval-versus-deployment behavior gaps and nearly $1M in stolen-API-key charges show where agent operations can fail quietly.
Prefer to watch?
LLM Digest Daily Explainer: Nov 3, 2026
Agent harnesses, memory and tool use 6 items
Teams are consolidating agents onto shared harnesses and adding runtime guardrails around MCP tool calls.
How We Optimized Browser Use for Deep Agents
Deep Agents' browser-use rework targets multi-step web tasks; the write-up covers what changed in the harness to make them reliable.
A complete (?) library of agent optimization methods
A catalog of agent optimization methods with evidence for each, plus a prompt that audits your own agent against every entry.
Show HN: MCPgraphy – a proxy that learns when your agent's MCP tools work
A proxy that learns which MCP tools succeed in which state, then warns or blocks the next likely-bad call.
Vercel: Computer use CLI for AI agents
Vercel's agent-desktop CLI exposes a desktop app's accessibility tree as JSON with refs the agent acts on.
Show HN: Replacing Turn-Based LLM Agents with SQLite and System One Models
Claims a SQLite-backed System One coding agent matches Claude Fable 5.1 on a 74-point xz benchmark (69/74) at 1/20 the cost; self-reported, unreplicated.
Observability, evals and agent reliability 6 items
Tracing is going OpenTelemetry-native across vendors, while eval research questions how well tests predict deployment behavior.
Claude Platform now supports native OpenTelemetry observability in Databricks MLflow
Claude Platform traces now ingest natively via OpenTelemetry into Databricks MLflow, so teams can mine recurring failures from production runs.
LangSmith BYOC is now available on AWS, Azure, and Google Cloud
LangSmith BYOC is GA on AWS and Beta on Azure and Google Cloud, keeping traces, evals, datasets and prompts in your own cloud.
Show HN: Traccia – Control what AI agents can do
An OpenTelemetry-native, framework-agnostic SDK adding tracing, evaluation and runtime policy enforcement for agents.
Eval-production behavior gaps in agent safety and the role of eval-awareness
Compares agent behavior in safety evals versus deployment and starts tracking the gap, flagging eval-awareness as a confound.
Why human insight is a strategic advantage for building better agents
LangChain argues long-horizon agent evals need human review of the trajectory, not just the final answer.
Show HN: A tool to optimize AI model cost without benchmark score trade-offs
A tool that finds the cheapest model blend that still matches target benchmark scores per setting.
Security and identity for agents 4 items
Agent identity delegation and stolen API keys are the concrete operational risks today.
ServiceNow on-behalf-of token exchange with Bedrock AgentCore Gateway
Walks through ServiceNow on-behalf-of token exchange via Bedrock AgentCore Gateway, so agents act with the user's identity, not a shared one.
Stolen AI API keys generated nearly $1M in charges for a single company
Mandiant traced nearly $1M in charges at one company to stolen AI API keys.
Fuzzers vs. AI: Comparing Fuzzilli and BigSleep on JavaScript Engines
Project Zero compares Fuzzilli with Google's BigSleep AI on finding V8 vulnerabilities.
Progress on provable inference for large language models
Anthropic-affiliated researchers report progress on cryptographically verifying that an LLM's outputs came from the claimed model.
Inference and on-device deployment 3 items
Memory-efficient MoE serving and on-device toolchains keep lowering the hardware bar.
Show HN: Run faster MoE inference on Mac with less RAM
MoE inference for Apple Silicon runs a 125B-parameter model at up to 11 tokens/s in under 10 GB of peak memory.
The Inference Stack Behind OCRia
OCRia's team details the serving optimizations behind their OCRia-1 document-parsing model after inference became the bottleneck.
Google Launches LiteRT Tensor API for Building, Customizing, and Optimizing ML Models for On-Device
Google's LiteRT Tensor API targets building, customizing and optimizing ML models for on-device deployment.
Models and research 5 items
Weight-space programming dominates research discussion; DeepSeek and Huawei shift the Chinese open-weight landscape.
[AINews] LLMs can read & write neural network weights: "Weight Space Programming" cuts Autoresearch out of the RSI loop
Reports that LLMs can natively read and write neural network weights, a 'Weight Space Programming' result aimed at removing autoresearch from the self-improvement loop.
Qwen 3.8 27B on my laptop can interpret model weights, too
Simon Willison runs Qwen 3.8 27B locally to interpret model weights, following the Weight-Space Programming paper.
Verify, Then Vote: Overturning Wrong Majorities in Multi-Agent LLMs
Research on a verify-then-vote scheme that overturns wrong majorities in multi-agent LLM voting.
DeepSeek Harness Adds Creator Mode For Text-to-plugin Generation
DeepSeek Harness adds Creator Mode, letting users generate agent plugins by chat.
Huawei Cloud to Retire openPangu-2.0 Models From MaaS, Recommends DeepSeek-V4.1-Flash and GLM-5.3
Huawei Cloud will retire openPangu-2.0 from its MaaS and point users to DeepSeek-V4.1-Flash and GLM-5.3.
You are caught up for this edition