We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
🔥 TOP SIGNAL
Coding agents are moving into the codebase control plane. Cursor’s Origin is rolling out an early beta around repos, pull requests, code browsing, and GitHub sync; its agents can answer about code, make changes, update PRs, or push a branch from the same surface. Zed’s DeltaDB makes the complementary bet below the commit: every operation between commits gets a stable identity, and every change links back to the agent conversation that produced it. Cursor’s caveat is the important one—“agent-native features ship soon”—so evaluate this wave on traceability and workflow integration before autonomy claims.
⚡ TRY THIS
Route by objective, then read the trace. Agents on Rails puts Claude Opus 5 at 58/63 runs, GPT-5.6 Luna as the cheapest and fastest model at a 3.3-minute median task time, and GPT-5.6 Sol as the best overall combination; four newly added models—including Grok 4.6, GLM 5.3, Gemini 3.7 Flash, and Claude Opus 4.8—did not reach the top. Start with Opus for expensive or high-risk changes, Luna for cheap quick passes, and Sol as the general default, then run the same matrix on your own repository. Treat the benchmark as a prior, not a verdict: models used provider-default effort, each task had only three attempts, and the suite is one small app—roughly 21 observations with ±5 points of noise. The trace data also suggests a review heuristic, not a rule: Claude Fable 5 failures usually missed the files containing the fix, while GPT-5.6 failures often found the right files and implemented the fix incorrectly.
Compile repeated tool use into “muscle memory.” Swyx’s pattern is to periodically use a larger model to turn a recurring sequence of primitive tool calls into a deterministic compound tool that smaller models cannot easily break. Find one repeated multi-step operation, make the successful sequence callable as one tool, and route routine instances through it.
Put agent guardrails in the framework, not the prompt. In @poteto’s account, Cursor’s agents window is 99% React and the team moved away from Solid partly because agents produced accidentally tracked code that created performance problems. Their Dune desktop framework bans direct
useEffectand exposes it only through framework-provided hooks; copy the pattern by making unsafe lifecycle behavior structurally unavailable, rather than merely documenting a preference.Make the remote machine the agent’s computer. Viticci reports coordinating dozens of threads from iOS; Codex Remote’s voice mode dispatches to individual threads, loads desktop context and plugins/skills, and can reopen threads on-screen. He used it over AirPods to set up a Mac mini, a remote KVM, and a connected Fingerbot. Theo’s alternative, T3 Code, emphasizes project creation, multi-PC management, remote configuration, and open source.
📡 WHAT SHIPPED
Cursor Origin entered early beta. Synced repositories update in real time while GitHub remains the source of truth; PR comments and replies sync both ways. Vercel supplies PR preview deployments and merge-to-production, while Depot and Buildkite run existing GitHub Actions workflows. The rollout covers paid plans except enterprise organizations that opt out. Kent C. Dodds also released a Kody Koala package for interacting with Origin through its API.
Zed DeltaDB opened early access. It records every operation between commits with a stable identity, links code changes to the agent conversation that produced them, and makes mid-run branching and live teammate annotation part of the workflow.
Claude Code
/designis in research preview. Run/designin the CLI or Desktop to get editable artboards, choose and tweak one, then have Claude implement it.LangChain and AWS added AgentCore Payments. When a tool receives a paid-API
402, the middleware checks the session budget, signs the payment, retries, and records the purchase beside the reasoning that triggered it in LangSmith.Agents on Rails expanded its public comparison. The update added Grok 4.6, GLM 5.3, Gemini 3.7 Flash, and Claude Opus 4.8, and published traces covering commands, diffs, and verdicts.
Claude Code CLI cut p99 CPU use by 2×. The team attributes the gain to changing Bun’s garbage collector from a fixed timer to an idle-only trigger, avoiding mid-turn CPU theft.
Omarchy’s community plugin repository passed 300 plugins. Until automated security reviews and versioning arrive, its maintainer guidance is to treat plugin code like an npm package, RubyGems gem, or AUR package—not as trusted code.
🎬 GO DEEPER
Study the Agents on Rails raw runs. The public directories include the full trajectory, shipped patch, hidden-test checks and verifier log, plus reward, steps, tokens, cost, and wall-clock data in
result.json—enough to build a review and routing benchmark instead of trusting a scorecard.Read the DeltaDB design page. Focus on the “between commits” model: stable edit identity, line-to-conversation lookup, free mid-run branches, and teammates joining before a commit or push.
Watch Rronak’s continual-learning talk. Swyx’s hook is the practical post-training problem: why GRPO is insufficient for their setting, why they moved to on-policy data, and how they handle the issues that introduces.
Editorial take: The durable coding-agent edge is shifting from “which model types fastest?” to a controllable loop: route by measured task fit, constrain the application, and preserve a trace from conversation to edit to PR.