AI AGENTS

A thought experiment with the sci-fi removed: freeze the models at today's capability, drop the price to zero overnight, and work out what actually changes for a working developer. Less than you fear, more than you think, and not where you expect.

CodeNib turns repository context into a data-system problem. That is the right direction for Claude Code, Codex, Cursor, and every agent that keeps rediscovering the same repo.

Luna drops from $1/$6 to $0.20/$1.20 per million tokens, Terra from $2.50/$15 to $2/$12, and Sol gets a paid Fast mode. What the new floor means for agent economics, Codex quotas, and the competition.

MCP Apps shipped with the 2026-07-28 final spec - sandboxed interactive UIs for MCP servers. How they compare to standard tool calling and standalone web UIs, and when to use each approach.

The Agent Skills spec gave agents progressive disclosure in three tiers - name, SKILL.md, bundled files. What it did not give them is a graph. Skills that link to each other, and say when to follow the link, let an agent navigate knowledge instead of front-loading it. Here is the argument, the measurements from our own 36-skill repo, and what to change.

Belay, Claude Code built-in guards, Codex CLI sandboxing, and MCP proxy patterns compared - how to protect your system from destructive commands, secret leaks, and prompt injection in AI coding agents.

How Claude Code, Cursor, Codex, GitHub Copilot, Aider, and Windsurf handle permissions, sandboxing, credential protection, and prompt injection. A structured comparison for engineering teams evaluating agent security.

AREX and the July deep-search papers point to the next useful research-agent primitive: a ledger of claims, constraints, failed paths, and unresolved questions that survives beyond the chat transcript.

Anthropic cut 80% of Claude Code's system prompt for Opus 5 and Fable 5 with zero regression on coding evals. The post landed on HN with 197 points and 133 comments. Here is what the article says, what HN thinks, and what it means for your agent harness.

Anthropic removed over 80% of Claude Code's system prompt for Claude 5 models. Here is how the rules changed and what it means for your CLAUDE.md files, skills, and system prompts.

Agents that critique their own output, learn from mistakes, and get better over time - the three patterns that actually ship, from simple reflection loops to tree search and meta agents.

Replit Agent 4 adds an infinite design canvas, parallel agents, and team collaboration to the prompt-to-app platform. Here is what changed, what it costs, and when to use it.

A practical comparison of the four authentication platforms developers reach for when connecting AI agents to third-party APIs: Arcade, Composio, Nango, and Stytch. OAuth 2.1, MCP support, integration counts, and which to pick by workload.

The DataFlow-Harness paper is a useful reminder that coding agents should not just emit scripts. For data work, the durable artifact is an editable, validated pipeline.

SearchOS turns web research from a growing chat transcript into shared state: frontier tasks, evidence graphs, coverage maps, and failure memory. That is the pattern serious deep-research agents need.

Cursor's latest agent-swarm experiment rebuilt a SQLite-like database from documentation and passed a held-out conformance suite. The bigger story is the shift from assigning code tasks to specifying, measuring, and governing a goal.

SWE-Pruner Pro points at a practical coding-agent design shift: do not only compress prompts outside the model. Teach the runtime to prune tool outputs before they become the next turn's context.

Microsoft's Resource2Skill paper points at the next agent-skills problem: converting videos, repos, articles, and reference artifacts into executable skills without losing provenance.

Prompt injection, sandbox escapes, and hallucinated dependencies are now documented, patched, CVE-numbered realities. Here is the threat model for agent-written code and the defenses worth adopting this week, ranked by effort.

A step-by-step guide to configuring an isolated Mac that Claude Code can fully control remotely - from SSH and Dispatch to phone-based control with Remote Control.