📚 ShroomDog Picks

When learning a new tool with Codex, the worst move is asking it to give you a lecture. A better pattern is to ask it for an entry point, a rough map, a tiny exercise, a teach-back check, and breadcrumbs for next time.

Anthropic explains how claude.ai, Claude Code, and Claude Cowork contain agents: model defenses miss, permission prompts create fatigue, and the hard boundary is the VM, sandbox, filesystem policy, and egress control.

Google Engineering Practices frames code review as code-health work, not a perfection ritual: approve CLs that improve the system, while aligning design, tests, speed, comments, and author habits around maintainability.

Codex is expanding from a coding assistant into a durable system for computer work: persistent threads, voice, steering, queuing, browser and desktop tools, automations, side-panel review, and shared memory all pull work from instruction toward execution and artifact review.

Nous Research proposes CNA, a method that uses contrastive prompts to find a tiny set of MLP neurons tied to refusal behavior. The interesting point is not just jailbreaks, but what this says about alignment fine-tuning.

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 SP-192, SP-197, and SP-207.

OpenAI and Anthropic both pushed /goal-like ideas into coding agents. A goal helps, but production agents also need strategy, constraints, health metrics, autonomy boundaries, and stop rules.

Whether Claude Code works inside a large codebase is not just about model scores. The real question is whether the team has built rails for the agent: maps, automation, on-demand tools, symbol navigation, internal-system access, and someone to maintain the whole operating setup.

Addy Osmani warns that default AI coding workflows help people close tasks, but do not automatically make them sharper. The difference is not whether engineers use AI; it is whether they use it to test and grow their own mental models.

Peter Steinberger says OpenClaw often runs about a hundred Codex instances in the cloud. The point is not showing off AI spend. It is testing what software work looks like when review, triage, security, reproduction, benchmarks, and meeting follow-up become always-on agent work.

Mitchell Hashimoto's point about Bun moving from Zig to Rust is not that Rust won and Zig lost. The more useful lesson is that programming languages are becoming more replaceable, and developer-tool companies need to manage technical narratives before the internet turns them into faction wars.

Anthropic lays out two 2028 scenarios for AI leadership: the US and its allies preserve their compute and model lead, or a CCP-controlled AI ecosystem catches up near the frontier. The essay centers on compute, export controls, model distillation, and whether democracies can set the rules first.

A practical agent engineering guide covering control loops, harnesses, context engineering, tool design, memory, multi-agent systems, evals, tracing, and safety boundaries.

Mem0 breaks down Codex CLI memory: not a vector database, but local Markdown, background summaries, credential scrubbing, and grep search. This post looks at when local notes are enough, and when a semantic memory layer makes sense.

Voice agents cannot reuse text-agent memory architectures as-is. Manthan Gupta breaks down why latency budgets, noisy transcripts, and cold-start identity make voice memory a different problem.

Garry Tan argues the real breakthrough in AI coding is not speed. It's turning tests, docs, and evals into a forward-only quality ratchet, so every change locks in what the team learned and makes the codebase harder to silently degrade.

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.

Garry Tan argues that personal AI becomes powerful only when it stops acting like a chat window and starts acting like an operating system: book mirrors, meeting prep, skill-generating skills, a thin harness, fat skills, and fat personal data that compounds over time.

Yage AI argues that OpenAI and Cursor are both moving from Skills toward Plugins, but for different reasons: OpenAI is building an execution-layer moat, while Cursor is building an editor-workflow moat. This gu-log rewrite explains why Skills create value but often fail to capture it.

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.