Fable 5 Redeployed, Agent Radio, 398-Subagent Audits — AI Daily Jul 01

904 messages · 89 active members

904

messages

active members

@Wootbro, @jcartu, @thewildzeno

top contributors

Overview

The day was dominated by Anthropic's redeployment of Claude Fable 5, which returned worldwide with US export restrictions lifted and inclusion in MAX plans for up to 50% of weekly usage through July 7 before shifting to API/credits only. Builders are skeptical of the original 'too dangerous' framing given Anthropic's own admission that Opus 4.8, GPT-5.5, and Kimi K2.7 can find the same vulnerabilities. Early qualitative reads are strong: Jason reported iteration counts on hard problems dropping from 20-300 to 10-30, and Samar ran a 398-subagent Fable harness against a 100k+ line codebase that surfaced 100+ bugs in a single pass. On tooling, @thewildzeno released Agent Radio — a lightweight local bus/router plus tmux-based persistent multi-agent setup that lets Codex, Claude, and other CLI agents communicate bidirectionally across repos, folders, and harnesses with zero token overhead for messaging. This solves the multi-repo handoff problem (backend↔frontend↔copy↔video) and beats ephemeral orchestrator subagents when long-running context matters. Related threads covered Claude Code config (--dangerously-skip-permissions, shift+tab, CLAUDE.md with @AGENTS.md inclusion), cross-harness skill portability, and Codex CLI's /import for migrating Claude setups. Sonnet 5 drew mixed reactions — fast for implementation after Opus-led planning, but token-hungry and prone to straying without tight guidance. A lively debate emerged on optimal orchestration (Opus plans → Fable critiques → Opus validates → Fable reviews), while GLM 5.2 is gaining traction as a cheap orchestrator with Fable reserved for high-value worker tasks. OpenAI teased GPT-5.6 sol, Opus 4.8 threw system-wide 529 errors, and builders discussed automating Meta ad rejection handling via the Marketing API.

Topics

Fable 5 Redeployed with 50% Weekly Usage Cap105 msgs

Anthropic brought Fable 5 back worldwide with export restrictions lifted, included in MAX plans for up to 50% of weekly usage through July 7 before moving to API/credits only. Members are skeptical of the safety narrative after Anthropic's own report noted Opus 4.8, GPT-5.5, and Kimi K2.7 can identify the same vulnerabilities. Heavy users are burning caps on deep codebase reviews and security audits before the window closes.

Fable 5 Real-World Wins and Multi-Agent Audits67 msgs

Builders reported major productivity leaps — Jason estimates iteration counts on hard problems dropping from 20-300 to 10-30, with landing pages, textures, and responsive design 'head and shoulders' above Codex or Opus. Samar ran a 398-subagent Fable-powered harness on a 100k+ line codebase, surfacing 100+ bugs in one pass. Tidemid ran a 6-hour /goal session where Fable started building logical extensions of the project unprompted.

Agent Radio: Cross-Harness Bus + Tmux Persistent Agents47 msgs

@thewildzeno released Agent Radio, a local router letting Codex, Claude, and other CLI agents communicate bidirectionally across repos and harnesses with zero token overhead for messaging. Paired with a tmux-based persistent multi-agent setup where frontend, backend, copy, and video agents live in separate sessions and hand off between roles — beating ephemeral orchestrator subagents when long-running context matters. Workspace presets spawn many terminal sessions at once.

Sonnet 5 Reactions and Claude Code Config45 msgs

Sonnet 5 landed with mixed feedback — fast for implementation after Opus-led planning, but token-hungry and prone to straying without tight guidance. Not worth it vs Opus for planning or Kimi/GLM for cost-sensitive work. New CC users learned to launch with --dangerously-skip-permissions, use shift+tab to cycle plan/auto modes, and adopt CLAUDE.md with @AGENTS.md inclusion for deterministic autoloading over flaky fallbacks.

Opus vs Fable Orchestration and Chinese Model Fallbacks32 msgs

Active debate on whether to plan with Opus and build with Fable, or the reverse. Zeno proposed an adversarial loop (Opus plans → Fable critiques → Opus validates → Fable reviews). GLM 5.2 is gaining traction as a cheap orchestrator with Fable reserved for high-value worker tasks, though a Coinbase report cutting AI spend 50% due to Chinese-model legal risk raised concerns. Opus 4.8 threw system-wide 529 errors, and OpenAI's GPT-5.6 sol preview may pressure Anthropic on pricing.

Key Takeaways

Fable 5 returns with 50% weekly limit inclusion through July 7 — front-load deep code reviews, security audits, and hard reasoning tasks before it moves to API/credits only.
Multi-agent Fable harnesses with 300-400 subagents are viable for full-codebase audits — Samar surfaced 100+ bugs in a single pass on a 100k+ line codebase.
Agent Radio enables cross-repo, cross-harness agent handoffs with zero token overhead for messaging; tmux persistent sessions beat ephemeral subagents for long-running context.
Launch Claude Code with --dangerously-skip-permissions, use shift+tab for auto mode, and use CLAUDE.md with @AGENTS.md inclusion for deterministic autoloading.
Emerging orchestration pattern: Opus plans → Fable adversarial review → Opus validates → Fable final check. GLM 5.2 works as a cheap orchestrator with Fable as premium worker.