ai-agents

A paper studying when AI agents should ask a shared human for clarification versus searching autonomously, proposing the Pairwise Coordination Policy (PCP) that achieves near-optimal costs (0.07% mean gap) on 864 synthetic cases and the lowest cost in an LLM agent case study.

Sonar CEO Tariq Shakir’s talk points out that AI agents have caused code output to surge (+2,450% in lines of code, +240% in commits), yet releases have grown only 30% — the “scaling from production to delivery” problem remains unsolved. He proposes a “code confidence formula” built by multiplying intent, guardrails, verification, and evidence, helping software factories deliver confidence rather than blind trust.

October roundup comparing CLI agent updates from OpenAI (Codex CLI), Google (Gemini CLI), GitHub (Copilot CLI) and open-source agents, covering local model support, pricing, licensing, and default telemetry behavior.

Kody (kody.codes) announced its open-source 'cloud for your team's agents' — a shared layer providing memory, packages, triggers, credentials, and apps that any MCP-speaking agent (Claude Code, Cursor, ChatGPT, Codex, etc.) can plug into without rebuilding setup each session.

Elon Musk shared a tip from Jason Ginsberg suggesting users treat Grok Bot like a teammate by giving it ambitious tasks typically assigned to a new hire.

The author compares agent guardrail approaches deployed at different layers this week — traffic proxies, browser secret-scrubbing, cryptographic agent identity for MCP, and GitHub Actions audit logs — and argues tool-call gating in the harness catches irreversible actions that wire-level proxies miss. Sharing their own Claude Agent SDK runtime 'Mainsheet' with a signed-receipt hash chain, they ask practitioners where enforcement should live and what each layer catches that the other misses.

Lantern launches People Infrastructure for Agents, a CLI-native layer that gives AI agents a single connection to 70+ people/company data and outreach tools so they can research and contact people via LinkedIn, email, and iMessage.

A practitioner asks whether AI agents can handle large legacy migrations (Java 8 and outdated Spring upgrades) at insurers or banks, weighing agent autonomy against bank-grade change-control processes and highlighting Orq AI's scoped-workflow software factory as a candidate approach.

A new Meta paper shows AI agents learn better from past attempts when experiences are routed into per-problem-type memory clusters rather than a single shared history, boosting performance by 15-30% of the max score while cutting API costs by 65-72%.

OpenAI announced a new internal benchmark evaluating how well its AI handles everyday human tasks like shopping and meal planning, reporting that GPT-6 Astra scored 93.1% on finding a kindergarten after a move and 90.4% on creating a diabetes-friendly meal plan.

A reflective blog essay about a lead developer whose role has shifted from writing code to supervising an AI agent, exploring a production bug caused by server/warehouse clock desynchronization and what 'checking' work means in the age of agents.

YC F26 startup Cyclon is building an AI phone experience where users can simply ask for tasks to be done, including writing, testing, and installing new apps on demand, letting AI agents operate a smartphone like a human user.

The post argues Muse for SMBs is an underrated consumer agent story, highlighting Meta's 3.64B daily active users, 500M+ SMBs, connector ecosystem plans, Shopify integration, and 25% of Muse users completing purchases averaging over $200.

Scrawl is a Product Hunt launch that lets you show your AI agent a problem using a visual frame, helping communicate issues visually to coding agents.

A 216-run benchmark compares Claude Code (Opus 5.5) and Codex (GPT 6 Astra) across agent-browser, Playwright, and Chrome DevTools via CLI, skill, and MCP on a browser task; all runs passed, tool choice depended on the agent, and a generated Python script vastly outperformed every agent run in speed and cost.

Phil Schmid explains how AI agents are collapsing their tool catalogs into a single `exec` tool — Code Mode — where models write TypeScript/Python scripts in a sandbox to chain actions, and argues MCP will return strong in 2027, with a minimal example for Gemini 4 Argon.

Suhail is hiring an AI agent engineer in SF to build open multi-agent systems (swarms, security agents, etc.) in a 0-to-1 role that blends applied AI research with product engineering; interviews begin this week.

OpenBMB's ForgeKernel agent autonomously optimized a GPU kernel against NVIDIA's FlashMLA black-box baseline over 22 hours, achieving a 9.8% speedup on B200, and is open-sourcing 27 optimized operators along with the ForgeKernel harness.

The author argues that human approval queues for AI agents aren't overhead but the key training signal: teams that capture rejection reasons and let actions auto-graduate shrink their queues, while teams that disable them see their agents fail. Cites Gravitee's 2026 report that only 14% of teams ship agents with full security/IT approval.

Microsoft distinguished engineer Stephen Toub recounts how he single-handedly ported the 400,000-line GitHub Copilot runtime from TypeScript to Rust in about 14.5 weeks using AI coding agents, at a cost of roughly 136 billion tokens and $120,000, discussing language choice, rollout strategy, and regression prevention.