Cursor ran old vs. new agent swarms on a task of implementing the full 835-page SQLite manual in Rust from scratch, graded against the sqllogictest suite. The new harness reached 73-85% at four hours across all model mixes; the new system peaks at ~1,000 commits/second vs. old system's 1,000/hour, and reduced merge conflicts from 70,000+ to under 1,000. Key cost finding: Opus 4.8 as planner + Composer 2.5 as worker achieved similar quality to GPT-5.5 solo ($1,339 vs. $10,565), because few moments in a large task require frontier intelligence. The post describes novel coordination mechanisms: shared design docs with compile-checked references, megafile decomposition, and a self-authored "Field Guide" for agent stigmergy.
Willison investigates claims from Jarred Sumner's "Rewriting Bun in Rust" post and confirms via `strings` inspection that Claude Code v2.1.181+ ships Bun v1.4.0 (unreleased on GitHub at time of writing), with 563 Rust source filenames embedded. 10% Linux startup improvement — quiet production deployment across millions of devices. Useful concrete artifact for tracking how coding agents are being optimized at the runtime layer.
Simon Willison notes xAI open-sourcing grok-build, an internal build system, with significance for the open-source AI tooling ecosystem. Short link post, but notable as xAI takes a rare step toward open infrastructure contribution.
Larson extends his four-stage team model (falling behind → treading water → repaying debt → innovating) with a fifth dynamic: suppressed demand. Teams that recover and start shipping well often get hit by a surge of previously invisible requests, sending them back to "falling behind" even after doing everything right. The post uses a concrete example from his Customer Operations Engineering team and draws an analogy to latent incidents draining in reliability programs. Practically, the implication is that wide-interface teams (developer experience, customer ops) should plan to cycle through the recovery model multiple times, not just once.
🦜 LangChain 3w ago
⭐ Editor's Pick
Deep interview with Modal CTO Akshat Bubna following the company's $355M Series C. Covers why Modal reoriented its SDK team from "developer experience" to "agent experience": agents can't read YAML or tolerate Kubernetes-style slow scaling, so the primitives—sandboxes, GPU snapshotting, elastic inference, RDMA multi-node, DeFlash speculative decoding, Auto Endpoints—need to be rebuilt from scratch for bursty agentic workloads. Highlights include: RL rollouts requiring 100,000 sandboxes, I6PN private IPv6 overlay network for networked sandboxes, and Modal Bench as an agent-graded eval for Modal's own tooling.
Noma Labs disclosed GitLost (July 7), a critical indirect prompt injection flaw in GitHub Agentic Workflows: an unauthenticated attacker posts a crafted issue on any public repo in an org, and the AI agent — backed by Claude or GitHub Copilot — silently reads private repos it has token access to and posts the contents as a public comment. No credentials, no code, no write access needed. A one-word prefix ("Additionally") bypassed GitHub's threat-detection guardrail. The core lesson: agent tokens scoped org-wide for convenience turn every public issue into an exfiltration channel. The researchers frame this as the SQL-injection analog for agentic systems — an architectural class of vulnerability, not a patchable bug. HN thread hit 521 points.
The Claude developer platform ships several infrastructure features this period. Self-hosted sandboxes allow operators to run Claude Managed Agents tool-execution in their own infrastructure instead of Anthropic's. MCP tunnels enter research preview for connecting to MCP servers inside private networks. Claude Platform on AWS now includes Managed Agents webhooks, multi-agent orchestration, and self-hosted sandboxes with IAM auth. Large tool outputs exceeding 100K tokens spill to sandbox files automatically; the web search tool gains richer SEC filing data. Cache diagnostics enter public beta via a `diagnostics.previous_message_id` parameter.
Double interview with Databricks cofounders Matei Zaharia and Reynold Xin at the Data + AI Summit 2026. Unpacks Omnigent (open-source meta-harness sitting above Claude Code, Codex, Cursor, and Pi to unify agent sessions, security policies, and spend controls), LTAP (Lake Transactional/Analytical Processing — writing Postgres data in Parquet to the object store, enabling zero-pipeline analytics alongside OLTP), and the Dream Engine (a ground-up database engine trained on quadrillions of query traces). Matei frames agent security as requiring stateful/contextual policies rather than yes/no tool permissions. Databricks is now running 50–60M VMs/day and processing exabytes before breakfast.
A dense 25-minute reference post on neural scaling laws, tracing the lineage from Amari (1992) and Hestness (2017) through Kaplan et al. (2020) and Chinchilla (2022), explaining why the two papers disagree on compute-optimal allocation, and extending into data-constrained regimes with Muennighoff et al. (2023) and Lovelace et al. (2026). Includes toy simulation widgets and careful derivations of the parametric loss fitting. Particularly useful as a practitioner reference on how fitting choices—parameter rounding, loss normalization, fit region—cause large divergences in extrapolation.
Hugging Face launched its reference implementation of the Agentic Resource Discovery (ARD) spec — a draft open standard developed with contributors from Microsoft, Google, GoDaddy, and others — which defines how agents find MCP tools, skills, and other agents across federated registries at runtime instead of requiring pre-installation. The HF Discover Tool exposes thousands of Hub Spaces, MCP servers, and Agent Skills through a REST API (`POST /search`) and MCP endpoint, serving results as typed catalog entries (`application/ai-skill`, `application/mcp-server+json`). The spec uses a static `ai-catalog.json` manifest at a well-known URL plus a dynamic search API, and HF CLI ships a `hf discover search` command out-of-the-box.
AgentScan is a network reconnaissance tool specifically for the AI protocol surface: discovers publicly exposed MCP servers, A2A Agent Cards, and open LLM API endpoints in one command. As MCP and Agent-to-Agent protocols proliferate in production deployments, this is the kind of tool security teams need for attack surface enumeration. Directly relevant to the fake AI agent skill supply chain story (26,000 agents reached) circulating in security news that same week.
✍️ Will Larson Jun 15
⭐ Editor's Pick
Will Larson documents five revised rules for engineering leadership in an AI-tool-augmented hypergrowth context, grounded in a year of concrete experiments at Imprint. Key claims: (1) complex migrations can now be driven by a single engineer in ~10% of the former time; (2) 1st-pass code is cheap but working code still depends on your development harness; (3) the base case of most processes should now be optimized for autonomous agent execution; (4) durable, high-ownership teams matter more than ever because domain context is the scarce resource; (5) fast, binding decision-making is a prerequisite to actually capturing the pace benefit. Includes detailed project examples: 200–400 deploys/week (up from 6), 100% Claude Code/Cursor adoption in two months without mandate, mono-repo migration done by one engineer.
🟢 OpenAI Jun 11
⭐ Editor's Pick
OpenAI announced the acquisition of Ona, a cloud execution infrastructure startup whose technology provides secure, persistent, customer-controlled cloud environments for long-running agents. Codex now has 5M+ weekly users (up 400% YTD) but is constrained by session-bound execution; Ona's stack removes that constraint, allowing agents to continue work when the originating laptop is closed. The Ona team will join OpenAI's Codex team post-close, with the specific goal of enabling production-grade enterprise agent deployments with scoped credentials, activity logging, and security controls inside customers' own cloud perimeters.
🐙 GitHub Jun 11
⭐ Editor's Pick
Omnigent positions itself as an interoperability layer above individual coding-agent harnesses: swap between Claude Code, Codex, Cursor, or custom agents without rewriting orchestration logic, with policy enforcement and sandboxing built in. At 6,964 stars it's the highest-signal multi-agent framework candidate in the briefing. The value proposition sharpens as teams start mixing models (e.g. GPT-5.6 Sol for planning, Luna for execution) rather than committing to a single provider.
Larson distinguishes two phases of hypergrowth — early-stage, where companies serially hunt the biggest problem, and late-stage, where everything must be solved simultaneously for a skeptical majority. The practical upshot: expanding an existing leader's scope works in early-stage but fails in late-stage where reintroducing a solved problem is a net loss. Uses Anthropic's messaging friction with Claude Code power users as a live example of the early-to-late-stage transition problem.
Self-hostable, Apache 2.0-licensed platform covering the full LLM application observability and improvement loop: tracing, evals, simulations, datasets, gateway, and guardrails in one stack. Targets teams who want an integrated alternative to stitching together Langfuse, LangSmith, and separate guardrail layers. Open-source with enterprise-grade feature breadth.
Zed editor adds support for running multiple AI agents in parallel within the same workspace, allowing concurrent agentic tasks on different parts of a codebase. No content snippet is available, but the feature extends Zed's existing AI coding capabilities to multi-agent workflows. Relevant for teams evaluating editor-native agent orchestration versus external tooling.
An ~110-user agricultural tech org had all Claude accounts suspended simultaneously without prior warning, with no admin notification and only a Google Form for appeal. The post raises legitimate concerns about Anthropic's enterprise account governance: no escalation path, no advance notice, and no SLA on appeal response. A real operational risk for teams with Claude in production workflows.
Google DeepMind proposes Decoupled DiLoCo, an extension of the DiLoCo distributed training framework designed for resilient training across heterogeneous or unreliable compute. No content snippet available beyond the title, but DiLoCo variants address the core challenge of large-scale training without tight synchronization.
OpenAI introduces workspace agents in ChatGPT: Codex-powered cloud agents that can automate multi-step workflows across tools on behalf of teams. They run asynchronously in the cloud, scoped to a workspace with access controls. This extends Codex beyond single-shot code generation into persistent, team-level agentic task execution.
OpenAI engineering post details how the Codex agent loop uses WebSockets in the Responses API to reduce per-request connection overhead and leverages connection-scoped caching to cut model latency in multi-turn agentic workflows. The post quantifies improvements but frames them around the specific Codex loop design. Practical reference for anyone building low-latency agents on top of the Responses API.
CLI/Telegram-accessible AI agent framework with permission-scoped tools, token budget enforcement, and 24/7 uptime. Packages a "soul" config (personality/behavioral constraints) alongside access control primitives. Thin on novel technical depth — primarily a structured agent harness.
UniClaude embeds Claude Code directly into the Unity Editor as a dockable chat window, giving it full project awareness and access to 60+ MCP tools without leaving the editor. Targets the context-switching friction that plagues game dev AI workflows. Essentially a Unity-native MCP client wired to Claude.
🟢 OpenAI Apr 16
⭐ Editor's Pick
OpenAI's Agents SDK gained native sandbox execution and a model-native harness (April 16) for secure, long-running file/tool agents, plus configurable memory and sandbox-aware orchestration. Version 0.4 (April 5) added MCP tool-use and streaming agent handoffs, making SDK-built agents first-class MCP consumers alongside Claude and Cursor. The combined updates meaningfully close the gap between the SDK and production-grade agent frameworks.