Coding Agents & Model Benchmarks

• GitHub Copilot agentic harness eval — GitHub publishes resolution rate + token efficiency data across 20+ models; useful baseline for comparing harness architectures.
• Ornith-1.0 open-source agentic coding LLMs — 9B–397B MoE family hitting 82.4 on SWE-Bench verified and 77.5 on Terminal-Bench; strongest open-source challenger yet for agentic coding tasks.
• Copilot code review now uses file exploration tools — Copilot PR review switched to the same grep/search tools as the CLI agent, cutting cost with no workflow change; signals convergence of review and coding agent infrastructure.
• levelsio on why coding agents need search/grep — concise articulation of why file-search capability is the hard requirement that separates agents from chat.
• OpenAI Codex economic research: agents shift unit of work — paper measuring Codex at frontier; frames shift from single-interaction to delegated long-horizon tasks as the economic inflection point.

Context & Model Strategy

• Context engineering as the agent moat — context window = RAM framing; the argument that prompt engineering optimises chatbots but context engineering is what keeps multi-decision agents reliable.
• Fable as planner/orchestrator, not coder — practical flow: use capable models to implement, send finished work to Fable for review; aligns with the cost-aware pipeline principle of routing by consequence.
• Where is Fable? (Theo) — community impatience signal; Fable's availability gap is becoming a talking point, worth watching for release.
• Claude Code features deep-dive (24 min) — Anthropic engineers walk through lesser-known Claude Code capabilities; worth queuing for feature audit.

Agentic SaaS & Tooling

• GitHub Copilot for Jira GA — generally available with model selection; Copilot now spans IDE → PR review → issue tracker, tightening the GitHub ecosystem lock-in.
• AWS Agent Toolkit for AWS — official MCP servers + plugins for Claude Code, Codex, Cursor, Kiro; lets coding agents build/deploy on AWS with guardrails baked in.
• T3code adds SuperGrok/X subscription support — xAI subscription now usable inside a coding agent IDE; expands Grok's distribution beyond chat.
• MagicPath as Cursor plugin — infinite canvas for interactive code inside Cursor; designer replaced Figma with it for several days (adjacent signal below).
• Vibe-coding displacing Figma for designers — concrete anecdote: designer using MagicPath + Cursor for clickable prototypes → direct engineer handoff; Figma's prototyping moat is leaking.
• Karpathy's "new paradigm" Slack bot debate — Theo's take on whether inline org-wide Claude access (vs siloed tools) is genuinely a new interaction model; worth watching for the infrastructure argument.
• Polygraph — cross-repo context + session memory for agents — addresses the multi-repo blind spot; agents maintaining memory across sessions is the unsolved problem this targets.
• Grass 2.0 — Claude Code from your phone — always-on mobile interface for coding agents; early signal for ambient/async agent workflows.
• BrowserAct — browser automation for AI agents — another entrant in the browser-agent layer; watch for differentiation vs Playwright-native approaches.
• Zaro — agents + apps from one prompt over your context — context-first app builder; pattern of "bring your own context, generate agent + app" is consolidating as a product category.

Vibe-Coding Resources

• claude-code-best-practice repo — community collection of subagents, commands, and skills for Claude Code; useful reference for harness patterns.
• Codex Orange Book (Chinese) — full-chain Codex usage guide with downloadable PDF; signals Codex adoption expanding in Chinese developer community.