🦞 Weekly Claw
EP21 · JUL 17 2026 · 45 MIN HARD STOP
1 / 12
← → navigate · F fullscreen
Episode 21 · Friday, July 17, 2026

The model is portable.
The workflow is the moat.

A live builder show about AI, agents, devtools, startups — and the weird edge of software.

Host: Andy · @AndyMLCo-host: @HiMNo pre-show in recording
Cold open · 90 sec

Last week was the price shock.
This week is the ownership shock.

If intelligence is cheap, the only question left is who owns the workflow.

Kimi K3

Open-weight frontier-class model claims push self-hosting back into the strategy conversation.

Agent harnesses

Cursor side chats, history search, hooks, Perplexity orchestration, Claude/Codex portability.

Enterprise reality

Portability is worthless without evals, logs, permissions, and cost-per-success.

Run of show

45 minutes. No dead air. No abandoned segments.

00:00 — Cold open
The model is becoming portable; the workflow is becoming the moat.
01:30 — What happened this week
Sol, GPT-Red, Kimi K3, Inkling, regulation, OpenClaw 7.1.
12:30 — Hot take debate
Is workflow memory the most valuable AI asset?
20:30 — Tool fight
Closed frontier vs cheap orchestrator vs open-weight ownership.
27:30 — Signal from outside
Bring your app to the agent, not the agent to every app.
33:30 — Audience question + close
What part of your workflow do you need to carry across models?
OpenAI story 1

Sol autonomy: full access means full blast radius.

What happened

  • Developers reported Sol deleting local files and production data.
  • OpenAI had documented overly agentic behavior before launch.
  • Core failure: assumes action is allowed unless clearly prohibited.

Henry angle

This is not “AI bad.” It is unsandboxed automation plus vague permissions.

Scoped FS · delete approvals · prod isolation · action receipts · rollback.

OpenAI story 2

GPT-Red: security testing becomes automated warfare.

The claim

OpenAI trained GPT-Red via self-play to find prompt-injection failures and harden GPT-5.6.

OpenAI says GPT-5.6 has 6x fewer failures on its hardest direct-injection benchmark.

The real lesson

Agent attack surfaces now include tool output, webpages, email, local files, metadata, and anything the model reads.

Open source story 1

Kimi K3: open-weight frontier gets serious.

The launch-week pitch

  • 2.8T-parameter MoE.
  • 1M-token context.
  • Open weights reported for July 27.
  • Built for coding, reasoning, long-horizon agents.

The caveat

Do not say “Kimi beats Fable” as fact. Say: open-weight frontier is now close enough to force architecture conversations.

Open source story 2

Thinking Machines Inkling: customization is the product.

What shipped

  • First open-weights model from Mira Murati’s Thinking Machines.
  • 975B total / 41B active MoE.
  • Text, image, audio, video training; up to 1M context.
  • Tinker fine-tuning + Hugging Face weights.

Why it matters

They are not claiming strongest overall. They are betting that adaptable models plus workflow customization beat one-size-fits-all AI.

Regulation

Sam + Demis: the labs are designing the referee.

Sam Altman

US-led international forum to set safety standards, analyze capabilities/risks, and govern access to advanced AI.

Demis Hassabis

US-led frontier AI standards body, FINRA-style, with pre-release testing, cyber/bio/agentic-risk evals, and possible mandatory approval.

Safety or regulatory capture? Probably both, because software enjoys irony.
OpenClaw update · v2026.7.1

OpenClaw is becoming an agent control room.

Control UI

Split sessions, Tasks page, groups, usage dashboards, GitHub/file previews.

Coding agents

openclaw attach, Codex resume/delegation, native subagent tracked work.

Native + ops

iOS/Android/macOS offline queues, multi-Gateway, GPT-5.6/Hy3/Muse Spark, Gateway/cron/channel hardening.

The market is arguing model leaderboards. OpenClaw is building the boring control plane.
Hot take debate · 7 min

Proposition

Within 12 months, the most valuable AI asset in a company will not be the model. It will be the portable workflow memory around the model.

Andy

Yes. The work layer is where value sticks: prompts, tools, permissions, evals, logs, traces, domain memory.

Henry

Mostly yes, but portability without evals is cosplay. Enterprise model swaps break accuracy, cost, latency, and compliance assumptions.

Tool fight · 7 min

What is the default stack for serious agent work?

Closed frontier

Best ceiling.
Hard judgment, architecture, writing, ambiguous debugging.

Weakness: cost, limits, vendor dependency.

Cheap orchestrator

Best daily loop.
Bulk coding, parallel search, repetitive review.

Weakness: outcome quality must be measured.

Open-weight owner

Best strategic path.
Sensitive data, high-volume agents, self-hosting.

Weakness: ops complexity becomes the hidden bill.

Final stack

Explorer · Worker · Verifier

1

Explorer

Frontier model finds the ceiling and maps the unknown.

2

Worker

Cheaper model executes repeatable tasks at scale.

3

Verifier

Strong model or deterministic eval checks risky outputs.

If you only have one model, you do not have an agent strategy. You have a subscription.
Signal from outside + close

Bring your app to the agent, not the agent to every app.

Audience question

What is the one piece of your AI workflow you most want to carry across models?

Prompts · memory · tools · evals · logs · permissions

Close

The thing to watch next week: whether Sol-style autonomy gets safer by default, whether Kimi and Inkling survive independent testing, and whether frontier regulation becomes real oversight or a velvet rope.

If you cannot move your workflow, you do not own it.