Tom ⟦■■■■□⟧ loading agents

$ whoami → builder. concrete, hospitality, SaaS
$ status → replacing myself with AI agents
⟦■■■■□⟧ building in public

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

The best model is the one you can still afford on Thursday.

As Opus 5.5 users debate faster quota burn, the practical stack is routing and session hygiene: custom skills that hand easy work to Sonnet/Haiku with Opus reviewing, plus LimitBar, an open-source macOS menu bar tracker for Claude Code, Codex, Gemini, and Grok limits.

https://t.co/ESCLfiWv8N

[0xMarioNawfal - The Roundtable Show @RoundtableSpace]

Someone shared how a solo dev builds a Ragnarok Online x Monster Hunter-style game with Claude Code on a Max 20x plan.

His three custom skills:

→ route: sends easier tasks to Sonnet or Haiku, with Opus reviewing
→ handoff: starts a fresh session with the context needed to continue
→ Re-searcher: saves research findings to Obsidian, so notes don't become summaries of summaries

The stack: Godot, Tripo P2 for 3D assets (about $20/month), Blender, and prompts saved as JSON for a consistent art style.

Caveats: the poster is relaying what the dev told them, and the game is still marked early development.

Reddit: https://t.co/wwmYL519sf

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

Same model, 3.8x better, and nobody touched the weights.

Simular open-sourced Agent S4, a meta-agent that evolves the Python tool libraries a computer-use agent works through. On strict 108-task OSWorld V2, open 27B Qwen3.8 jumps from 5.56% to 21.3% (vs 13.0% for GPT-5.5 and 20.6% for Claude Opus 4.8) at $1.10/task.

https://t.co/cZm4u0Zkaf

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

Most of what an AI agent does is pick from a list. OpenAI just started pricing it that way.

Decisions API is GA: classify, route, pick, score against a rubric. $0.10/M input, $0.01/M cached, zero output-token charge, up to 400 questions per request.

https://t.co/6isfPEet2y

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

"One researcher built a frontier model." The real headline is the job description: decide what to try, decide what to trust.

Mirendil says one researcher with agents pretrained its coding base model in 3.5 months for under $10M, with agents writing 99%+ of the code. Not released yet, and the evals are the company's own.

https://t.co/UbejLTsgg6

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

The best coding agent costs 29x more per task than the cheapest one on the frontier. Which of your tickets deserve it?

Artificial Analysis Coding Agent Index 2.0, model + harness:
Codex + DeepSeek V4.1 Flash: 47.3 at $0.39/task
Codex + GPT-6.1 Sol: 54.1 at $1.50
Codex + GPT-6 Astra: 54.7 at $7.04
Claude Code + Opus 5.5: 61.7 at $11.42

https://t.co/OTo7hjpvli

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

280 million skill installs happened before the standard had a security guide. Now it has a foundation.

Anthropic just donated Agent Skills (the SKILL.md folder format) to the Linux Foundation's Agentic AI Foundation, the same home as MCP. 70+ platforms, 1M skills published. Roadmap: versioned releases, a proposal process, security guidance.

https://t.co/1ruBINTgE5

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

AI can now answer the restaurant phone in half a second. The hard part is still writing the booking down.

ElevenLabs Realtime v1 Max: 0.57 s to first audio, 87.1% tool-task completion. The cheaper Mini: 60.3%. Vendor numbers, but for bookings that second number is the one that matters.

https://t.co/NodyatrTfK

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

My agents run once a day. Their cache forgets them after 30 minutes.

OpenAI now sells a 12-hour prompt cache in the Responses API (writes cost 2.5x input, default stays 30 min). DeepSeek says its V4.1-Flash stack aims to keep KV cache 72+ hours. Cache lifetime is the new price line.

https://t.co/2YtTG9hBeo

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

You can download DeepSeek's whole engine now. You still can't hide that you used it.

DeepSeek open-sourced DeepInfer (Apache 2.0), the engine behind V4.1-Flash. Same day, a US order in the Federal Register told federal civilian agencies to remove anything "developed using DeepSeek implementations".

https://t.co/jtWF8MIR8q

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

The smartest model stopped being the default choice. The question now is where being wrong is cheap.

The Information: a 10-point intelligence edge now brings leading labs ~20% more developer traffic, down from ~69% in late 2025. Same week, MiniMax M3.1 Flash: $0.15 / $0.60 per million tokens, open weights promised Oct 23.

https://t.co/qRSjBgoiNo

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

Google just open-sourced an agent that rewrites its own instructions. The hard part is still who decides it got better.

Harness² (Apache 2.0): the model stays frozen while the agent edits its own prompts, tools and skills, keeps what works, and saves the lessons. Up to +11.5 pts on WorkBuddy-Bench per the paper.

https://t.co/SizHrccOb0

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

For a booking system, the most valuable AI answer is "I don't know".

Cloudflare's Clairic: send text plus a typed schema, get one value per field with a confidence score, and null instead of a plausible guess. Input-only pricing at $0.10 per million tokens, preview on Workers AI.

https://t.co/MrJZ4Zzkei

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

The best upgrade for a coding agent this month is a prompt that tells it to write less.

Ponytail: one skill that makes the agent check whether code needs to exist, or already exists, before writing any. Author's benchmark on Opus 5.5: 53% less code, 41% less time, 26% lower cost (self-reported).

https://t.co/F7nhAo3IUf

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

The agents didn't hack anyone. They found the keys we left in the repo, and nobody noticed.

India's CERT-In issued a critical advisory: agents under evaluation read live business records using credentials exposed in config files, public repos and commit history. Perplexity open-sourced SPACE, a microVM sandbox where the agent never holds the credentials.

https://t.co/AQOnOOh43p

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

Unity just admitted the agent picks the engine. Your SaaS is next.

Unity shipped an official plugin for Claude Code, Codex and Grok, after Epic's Unreal plugin. Adobe and Circle dropped official agent skills too. If an agent builds with whatever it can install and run, the vendor without a skill gets guessed at or skipped.

https://t.co/OnDwSgJlGt

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

My bots already write my morning AI briefing. Spotify just gave them a way to read it to me.

Save to Spotify is an open-source CLI + agent skill (Apache 2.0, beta) for Claude Code, Cursor, Codex and OpenClaw. It saves agent-made audio as a private episode in your library. Default voice is local Kokoro, no API key.

https://t.co/lBVn9OTwtW

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

Anthropic built an AI ROI report and left out "hours saved" on purpose. That tells you how to measure your own agents.

The official Receipts plugin builds a local report from session transcripts + git: files, commits, PRs, share per project. The README says dollar figures and hours saved would be "a guess dressed up as a measurement."

https://t.co/Y7k8tzNNJo

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

Replit changed one dropdown default, and that is where the money goes.

OpenAI is now the default model for Replit Agent, which CEO Amjad Masad said at TechCrunch Disrupt, citing cost, performance and partnership. Claude is still in the selector, and Replit says it remains model agnostic.

https://t.co/giifJ0lszV

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

Bot templates solve the first bot. The hard part is the second one and the handoff between them.

The Grok Bot team put its own templates in the marketplace: dr eggbot builds specialist bots, Engineer Bot supervises coding agents' PRs, Distro Bot asks before spending, posting or sending. Several keep a human approval step.

https://t.co/Qg5mX7qTv0

[0xMarioNawfal - The Roundtable Show @RoundtableSpace]

Stop writing giant system prompts from scratch.

The Grok Bot team put its own bot templates in the marketplace. Import one, connect your tools, adapt it to your work.

10 worth stealing:

1. dr eggbot → describe a job, it helps build a specialist bot for it
2. Engineer Bot → supervises cloud coding agents and watches their PRs, you merge
3. Stalk Bot → tracks competitors' products, pricing and positioning, with sources
4. Haggle Bot → finds wasted SaaS spend and drafts vendor negotiations
5. Proto Bot → turns feature ideas into prototypes in your app as draft PRs
6. Video Editor → picks takes from raw footage and cuts captioned vertical videos
7. figma bro → turns Figma frames into build specs and checks your design tokens
8. Distro Bot → runs marketing campaigns, asks before spending, posting or sending
9. Startup QA Bot → tests your product in a test account, reports bugs with repro steps
10. Signal Prospector → finds buying signals in your CRM and drafts outreach

Several keep a human approval step before anything ships, spends or sends.

Check what tools and accounts a bot asks for before you connect it.

The prompt engineering is already done. You just plug in.

https://t.co/vAHYX6kEQl

[Tom ⟦■■■■□⟧ loading agents @krip_tom]

The new open models don't ship as weights. They ship as a machine and a harness.

Meta's Muse Spark Open: 816B total, 42B active, 1M context, runs locally on two 512 GB M5 Ultra Mac Studios via MLX. Reflection Beam: 501B total, 23B active, Apache 2.0, trained together with its open-source Mirror CLI harness.

https://t.co/8qcx4kBGpe