# agent

Letta open-sourced trajectory, a format that normalizes work records from Claude Code, Codex, OpenClaw, and other harnesses for agents to read. In sampled transcripts, it cut token counts to roughly one-fifth of native. Cross-harness learning starts with a standard experience format.

Agent evals must verify not only the response, but whether the world actually changed. Anthropic shows how to combine graders, isolate environments, and maintain an eval suite over time—and why 20–50 real failures are enough to start.

Claude Tag, Anthropic’s second-generation Claude in Slack, launched in June 2026. Each thread is a persistent work session anyone in the channel can steer. Its sandbox evaporates and never holds credentials—the security baseline for building our own.

Claude Tag officially supports only Slack. To get an on-call AI assistant on LINE, build one: run OpenClaw on a VPS, let its gateway receive messages from LINE’s cloud, and have the agent reply. Here’s the reusable skeleton, three security must-haves, and why freedom brings responsibility.

Jarred Sumner rewrote 535K lines of Zig into Rust with 64 Claude agents in parallel, adversarial code review, and mechanical porting. 11 days later: all tests green, memory leaks fixed, binary 20% smaller.

Phil Chen shares six years of career lessons — from his own startup through Helm AI, Scale AI, OpenAI, and Google: when agents can solve every well-defined problem, what stays valuable is finding problems, sprinting the last mile, and everything that cannot be graded by a loss function.

Open any agent skill and it's stuffed with 'be more detailed,' 'be thorough'—lines that look diligent but don't change the model's behavior at all. Matt Pocock names the no-op trap, plus how to spot a dead instruction versus one that actually pulls its weight.

Sakana ships Fugu: a multi-agent orchestration system behind one API, sold as "AI sovereignty." But a researcher who read the tech report tears it down — a closed orchestrator on closed models means you control less, not more, and it wins benchmarks while never reporting cost.

Most coding-agent best practices from six months ago are now out of date. The new playbook: bigger tasks, longer sessions, and adversarial review so the agent verifies its own work — the engineer just shovels coal into the engine.

Loop engineering is sold as designing orchestration and spinning up agents — but the tools now do that half for you. The half still hard, still deciding the result, is the verifier. Anthropic's Bun port is the tell: 99.8% of tests pass, yet the announcement says not yet in production.

Every two hours, Codex drafts email replies for review. The drafts are good — he edits them anyway. Those edits are context too, and most automations throw them away. The fix: an inner loop brings context to the work; an outer loop recovers context from the review diff.

A six-word phrase seized the AI-coding timeline, but nobody boosting it agreed what it meant. This is not the how-to; it is why the loop blew up, its five-year lineage, why the loop is now the costly part, and why the durable asset is the skill it calls.

Microsoft CEO Satya Nadella on the future of the firm in an AI economy: build two kinds of capital — human capital and token capital. The real moat isn't picking the strongest model, but a learning loop that compounds. Plus a warning: don't let a few models eat every industry.

Dimillian (an iOS dev now at OpenAI) wrote a field guide for Codex Mobile. The part worth keeping is a mental model that holds across tools: your phone is not a shrunken terminal, it is the control center that keeps you making decisions while the agent does the work.

Two of the most senior AI engineers alive said the same thing this week: stop prompting your agent, design loops that prompt it for you. Loop engineering unpacked — open vs closed loops, the six building blocks, prompt vs loop engineer. Plus: spotting one smooth ad sewn into the lesson.

Vaibhav Srivastav's Codex prompt is interesting because it describes an agent maintenance loop: look back at recent work, find repeated workflows, and package only high-confidence patterns into Skills, automations, or subagents. It is agent dreaming: turning busy work into capability.

OpenAI's Cookbook frames Codex Goals as a thread-scoped completion contract: the objective persists, but completion must be checked against evidence. This post fills in the official spec angle around GP-192, GP-197, and GP-207.

Codex `/goal` is not a wish machine. Chris Hayduk's real point is engineering discipline: give the agent a measurable finish line, a fast feedback loop, and Markdown files that work as long-term memory.

LLMs are not gods, and they are not just tools. They are more like DOTA teammates: great at last-hitting, occasionally great at feeding. The human job is not to fight AI for the same lane, but to cover taste, map awareness, context ownership, and strategic judgment.

Thariq explains why HTML is replacing Markdown in Claude Code workflows: not as prettier output, but as readable, operable, shareable artifacts that keep humans inside the agent decision loop.