Coding agents

Coding agents & models

• Fable 5 lands broad: from Jul 20, Fable 5 folds into Max/Team Premium at 50% limits; Pro/Team Standard keep credit-based access plus one-time $100 credit — direct signal on what model tier you get access to without add-on spend.
• Kimi K3 2.8T-A50B ships: largest open model release yet, benchmarked Opus 4.8-class quality at Sonnet 5 pricing — biggest open-weight threat to closed frontier pricing this cycle.
• Kimi K3 on ChatLLM auto-router: routes simple-coding to Grok 4.5, medium-agentic to Kimi K3, hard to Fable 5/Sol, usable inside Codex/Claude/Abacus — a working example of prompt-complexity-based model routing you could mirror in your own pipeline.
• Claude Code subagents go background-by-default: next Claude Code version runs subagents in background unless told foreground — matches the async-delegation pattern already in your workflow.
• Artifacts land on Pro/Max: Claude Code can now publish live self-contained artifacts to claude.ai on Pro/Max, not just higher tiers.
• levelsio ditches Claude Code for Kimi K3: anecdote citing Claude Code's Opus→Sonnet safety downgrades stalling a project for 2 weeks, Kimi K3 pushed through — useful anecdotal signal on safety-guardrail friction, weigh against vendor-neutral benchmarks.
• Run Kimi K3 inside Codex via OpenRouter: config snippet to point Codex's model backend at Kimi 3 through OpenRouter.
• Kimi Code console walkthroughs (alt link, alt2): repeated how-to threads on installing OpenCode via Claude Code/Codex, then connecting to Kimi Code ($19/mo) as backend — same jailbreak-the-harness pattern posted three times today.
• Theo on gpt-5.6/Kimi K3 inside Claude Code: video comparing what Claude Code gets right vs Codex, with his multi-model setup inside Claude Code's harness.
• claudex alias for Codex→Sol: CLIProxyAPI-based alias routes CLAUDE_CODE_SUBAGENT_MODEL to GPT-5.6-Sol from inside Claude Code, 5-minute setup.
• Kimi K3 launch on Product Hunt: billed as the world's first open 3T-class model.

Context & token engineering

• code-review-graph MCP/CLI: Tree-sitter-based local code intelligence graph, incremental change tracking, benchmarked context-token reduction for review and large-repo workflows — direct competitor/reference for your cost-aware-pipeline context strategy.
• TS-only CodeGraph, 90% token cut: author claims other codegraph tools actually spend more tokens and slow indexing; worth a skeptical read against code-review-graph above.
• OmniRoute free AI gateway: single endpoint fronting 231+ providers (50+ free) for Claude Code/Codex/Cursor/Cline/Copilot, claims RTK+Caveman-stacked compression saving 15-95% tokens with auto-fallback — compression claim overlaps directly with your caveman-mode token work, worth verifying independently.

Agent orchestration & tooling

• openai/codex-plugin-cc: official OpenAI plugin to call Codex from inside Claude Code for reviews (/codex:review, /codex:adversarial-review) or task delegation (/codex:rescue etc.) — this is the plugin your own orchestration rules already reference for peer-review with Codex.
• Orca multi-agent ADE: desktop/mobile app running Codex, Claude Code, OpenCode, Pi side-by-side each in its own git worktree, with usage/rate-limit tracking across accounts — same worktree-per-agent pattern your swarm setup uses, packaged as a GUI.
• herdr terminal agent multiplexer: detach/reattach-persistent multiplexer for multiple coding agents with a socket API so agents can spawn/read/wait on each other's panes — overlaps with tmux-based swarm orchestration you already run.
• SnapState: persistent agent state: positions itself as persistent state for AI agent workflows, thin on detail from the HN post alone.
• Vercel's Andrew Qu on agents as new software: argues agents are a distinct software category from Vercel's eve framework work, with skills/sandboxes/agent-readable websites as the emerging primitives.
• Autoresearch self-improving loops: Introspection co-founder on "autoresearch" outer loops where agents maintain/improve their own recipes, humans staying in the loop — conceptually close to your reflect/self-improving skill stack.

Agentic SaaS & voice

• OmniDesk v2.4.0 offline voice prompting: on-device, fully offline voice-to-Claude-Code-terminal prompting.

CLI/TUI & platform

• Copilot code review customization: adds a firewall, custom setup steps, independent runner configs, and reads custom review instructions from the head branch — GitHub closing the gap with configurable review agents like Codex/Claude review plugins above.

Commentary

• Claude Fable 5 and AI safety framing: critiques Anthropic's "silent model manipulation" of safety-flagged research queries, notes a mid-release fix to use a classifier consistently — relevant background if you've hit unexplained refusals/downgrades.