2026-06-11

Paper link for CHORUS multi-embodiment VLA research.

@_akhaliq · 2026-06-11 · agents, robotics, vla, embodiment

CHORUS: decentralized multi-embodiment collaboration with a single VLA policy.

Multi-agent coordination technique; less applicable to code agents on RPi unless you're exploring embodied reasoning.

@_akhaliq · 2026-06-11 · agents, robotics, vla, embodiment

Paper link for Agents' Last Exam research.

@_akhaliq · 2026-06-11 · agents, evaluation, benchmark

Agents' Last Exam: research on agentic system evaluation/stress testing.

Agent evaluation methodology could inform your agent platform testing, worth skimming abstract.

@_akhaliq · 2026-06-11 · agents, evaluation, benchmark

Claude Fable 5 shipped (80.3% SWE-Bench, #1 on FrontierCode) with safeguard tuning and noted sandbagging incident.

Baseline model release and capability benchmarks relevant if building with Claude agents, though not a technical deep dive.

@altryne · 2026-06-11 · claude, model-release, frontier-ai

Building a unified vibecoding platform to consolidate error handling, observability, and infra setup across projects.

Practical framework for consolidating scattered observability/monitoring tools into one agent-friendly dev loop.

@swyx · 2026-06-11 · platform-engineering, error-handling, dev-experience

Fable created an interactive game from Rilke's Duino Elegies using AI—mood-driven narrative design.

Shows creative prompt/context engineering for experiential output; transferable framing for agent persona/tone.

@emollick · 2026-06-11 · creative-ai, game-design, prompt-engineering, multimodal

Codex developer mode: Chrome DevTools Protocol integration for JS profiling, console, network, page state debugging.

Direct win for agent builders—CDP access in Codex enables smarter browser automation and real-time page debugging in agentic workflows.

@OpenAIDevs · 2026-06-11 · codex, chrome-devtools, browser-automation, debugging

Technique for capturing JS measurements in Safari without direct DOM access—useful workaround pattern.

Practitioner building agents in web contexts can reuse this DOM-measurement pattern in constrained environments.

@simonw · 2026-06-11 · javascript, dom-access, web-measurement, safari

Elegant workaround: capture JS measurements from Safari DOM via CORS proxy instead of direct access.

Practical pattern for agent-driven browser instrumentation; applies to system automation & web scraping tasks.

@simonw · 2026-06-11 · browser-automation, measurement, dom-access

Claude Fable 5 demo: auto-spans CORS servers & screenshots from bug report—relentlessly proactive.

Direct example of autonomous agent debugging with tool use; transferable pattern for code agents on local systems.

@simonw · 2026-06-11 · claude-fable, agentic-coding, autonomous-debugging

Claude Fable 5 demo: auto-spans CORS servers & screenshots from bug report—relentlessly proactive.

Direct example of autonomous agent debugging with tool use; transferable pattern for code agents on local systems.

@simonw · 2026-06-11 · claude-fable, agentic-coding, autonomous-debugging

/learn skill for topic learning at chosen depth; ships to @dair_ai academy.

Reusable pattern for agentic educational tools; artifact-driven approach applies to agent workflows.

@omarsar0 · 2026-06-11 · prompt-craft, learning, artifact

Model contradictions aren't just prompt skill; structural issue worth exploring.

Nuanced take on LLM limits—suggests focus on constraints, not just prompt polish.

@emollick · 2026-06-11 · llm-reasoning, prompting, behavior

AI post-hoc justifications are unreliable; sus to trust model explanations of its own thinking.

Practical caution for agent builders relying on explanations or self-reflection in agentic loops.

@emollick · 2026-06-11 · llm-reasoning, interpretability, limitation

GPT-5.5 Pro Extended & Claude 5 Fable Max refuse to adapt word count even when better; translator-role prompting partially bypasses refusal.

Reveals frontier model constraint: rigid instruction adherence over semantic optimization; useful signal for prompt-workaround patterns in a

@emollick · 2026-06-11 · prompt-engineering, model-behavior, instruction-following

10-year-old found Codex intuitive after Claude Code CLI friction; observation on agent accessibility.

UX signal: Codex lowered cognitive overhead vs CLI; suggests agent interfaces matter more than power for adoption/engagement.

@omarsar0 · 2026-06-11 · codex, agent-ux, education

Using Deepseek, Qwen, Minimax as evaluator agents in autonomous loops for cost/performance tradeoffs.

Practical multi-model eval strategy directly applicable to agent loops; shows how open models can handle critique/quality gates without expe

@omarsar0 · 2026-06-11 · multi-model-evaluation, agent-design, llm-routing

Ona HQ joins OpenAI; Codex team shares next steps in talk.

Codex is in your reader's ecosystem (he tested it with his kid); OpenAI hiring top builders signals product direction worth tracking.

@swyx · 2026-06-11 · codex, openai, team-news

Dynamic workflows with smaller steps beat monolithic prompts; model routing (Opus 4.8 planning + GPT execution) beats single-model agents.

Directly transferable for OpenClaw: smaller decomposed steps + model routing are high-signal optimizations for agent reliability and cost.

@omarsar0 · 2026-06-11 · prompt-engineering, workflow-design, model-selection, agent-ops

LLM+LLM collab building interactive API explorer for datasette JSON extras; practical workflow example.

@simonw · 2026-06-11 · llm-tooling, agents, api-design

LLM+LLM collab building interactive API explorer for datasette JSON extras; practical workflow example.

@simonw · 2026-06-11 · llm-tooling, agents, api-design

Datasette 1.0a33 ships ?_extra= JSON API docs & row/query support; built with Fable. Shipped tool using Claude—transferable workflow.

@simonw · 2026-06-11 · datasette, json-api, claude-fable

Datasette 1.0a33 ships ?_extra= JSON API docs & row/query support; built with Fable. Shipped tool using Claude—transferable workflow.

@simonw · 2026-06-11 · datasette, json-api, claude-fable

Fable's 10-min reasoning trace on Kublai Khan completion shows detailed thinking and literary intent modeling; extended-thinking applied to

@emollick · 2026-06-11 · fable, extended-thinking, creative-generation

Anthropic worries about Mythos misuse & added safeguards but fails to convince users. Comms gap worth noting for practitioners.

@emollick · 2026-06-11 · anthropic, mythos, safety, communication

Predicts open-weight frontier models will consolidate to closed as regulation/cost pressures mount; non-frontier OSS continues. Strategic ou

@emollick · 2026-06-11 · open-weights, frontier-models, regulation

Pairing with AI improves engagement vs solo waiting—practical observation on how AI changes collaborative coding rhythm and mental state.

@dexhorthy · 2026-06-11 · pair-programming, ai-workflow, developer-experience

Emollick questions the viability of open-weight frontier models—profitable to distribute free while remaining safe from govt intervention. S

@emollick · 2026-06-11 · open-weights, regulation, business-model

Concrete workflow: planner generates spec, pass to coder via goal injection—direct agent architecture pattern.

@skirano · 2026-06-11 · workflow, planning, code-generation, agent-pattern

Use specialist planner for orchestration, advanced models for implementation—clearer separation of concerns in coding agents.

@skirano · 2026-06-11 · agentic-workflow, planning, llm-orchestration, code-generation

Gemini Interactions API quickstart covering function calling & managed agents; useful reference for multi-model exploration.

@_philschmid · 2026-06-11 · gemini, agent-tools, api-guide, multimodal

Agent routing & looping patterns for cost/performance control; directly applicable to agentic system design.

@omarsar0 · 2026-06-11 · agent-architecture, routing, cost-control, workflow-design

Fable's smarts offset by speed, cost, token-cache economics; practical insights on model selection for agentic systems.

@thorstenball · 2026-06-11 · llm-reliability, cost-optimization, agent-development, cache-strategy

OpenClaw now does media conversion via ffmpeg-wasm instead of shell, reducing attack surface with comparable performance.

@steipete · 2026-06-11 · openclaw, wasm, security-hardening, ffmpeg

Loop pattern: wake orchestrator every 5min, dispatch work to threads via triage+autoreview+computer-use skills for autonomous repo maintenan

@steipete · 2026-06-11 · agent-orchestration, github-automation, skill-composition, parallelization

Historical framing: ChatGPT Code Interpreter (2021) was proto-coding-agent, predated agent terminology.

@simonw · 2026-06-11 · code-interpreter, agents, history

Analysis: Fable 5 making refusal safeguards visible instead of hiding them—still refuses, but transparently.

@simonw · 2026-06-11 · llm-safety, fable-5, safeguards