Computer Use Agents Get Cheap

Edition 159 | February 19, 2026

Love the idea of a reverse CAPTCHA: ‘Are you a robot?’ Please demonstrate. 🦞

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

In today’s issue…

• Claude Sonnet 4.6 makes reliable computer-using agents affordable

• Qwen 3.5 brings open-weight visual agents to real workflows

• METR finds where coding agents hit their productivity ceiling

• Manus and Lindy bring multi-step agents directly into Telegram and iMessage

…and more

🔥 INCASE YOU MISSED IT

Readers’ favorite items from the past week

1. Apple studied how users expect to interact with AI agents

2. ElevenLabs secures insurance for voice agents

3. OpenClaw (formerly ClawdBot/Moltbot) joins the OpenAI team

4. Airbnb says AI now handles ~33% of customer support in NA

5. Harvard Business Review argues companies now need Agent Managers

📌 THE THURSDAY BRIEFING

Claude Sonnet 4.6 with 1M context & "Computer Use" parity

📡 Signal: Anthropic just made Claude Sonnet 4.6 the new default model, effectively killing the trade-off between speed and power. For the first time, a mid-tier "Sonnet" model features a 1M token context window and reaches parity with the flagship Opus-class in agentic coding and Computer Use benchmarks (scoring 72.5% on OSWorld). It also introduces Context Compaction, a feature that automatically summarizes older parts of a conversation mid-task to prevent agents from hitting context limits during long-running projects.

For builders: The real story is cost. Until now, reliable Computer Use required expensive Opus-class models. Sonnet 4.6 brings that same "human-level" reliability (multi-step web forms, complex spreadsheets) at 1/5th the price. Adaptive Thinking lets the model decide when to "think longer" before acting, making it a more stable backbone for agents managing legacy systems without modern APIs.

Alibaba launches Qwen 3.5 with native Visual Agentic capabilities

📡 Signal: Alibaba released Qwen 3.5, a 397B-parameter open-weight model built for the "agentic AI era." The headline: native Visual Agentic skills, the model sees UI screenshots and executes multi-step workflows across mobile and desktop apps in a single inference pass. Using Mixture of Experts (only 17B active parameters), it's 60% cheaper and 8x more efficient than its predecessor. On OSWorld-Verified it scores 62.2%, competitive but behind Sonnet 4.6's 72.5%. The hosted Qwen 3.5-Plus adds a 1M token context window.

For builders: Alibaba is betting on capability per dollar. The native vision-language architecture skips expensive "screenshot → OCR → action" pipelines. The real differentiator: it's Apache 2.0 open-weight—self-host, fine-tune, no API lock-in. If Anthropic's Computer Use is the premium option, Qwen 3.5 is the self-hosted alternative: not quite as accurate, but a fraction of the cost with full infrastructure control.

🤝 WITH WISPR FLOW

Better prompts. Better AI output.

AI gets smarter when your input is complete. Wispr Flow helps you think out loud and capture full context by voice, then turns that speech into a clean, structured prompt you can paste into ChatGPT, Claude, or any assistant. No more chopping up thoughts into typed paragraphs. Preserve constraints, examples, edge cases, and tone by speaking them once. The result is faster iteration, more precise outputs, and less time re-prompting. Try Wispr Flow for AI or see a 30-second demo.

Start flowing free

 🤖 AGENT OF THE WEEK

👋 Welcome back to Agent of the Week!

If you're like me, you're in meetings all week. Client projects, internal builds, sponsor calls, random "quick syncs" that turn into 45-minute roadmap sessions.

The meetings aren't the problem. What happens after is.

Transcripts get saved, context gets scattered. A few days later I'm asking: What did we actually decide? Who owns that? What changed since last week?

So I built a Meeting Delta Agent using Claude Cowork.

If you don’t know what Cowork is, here's the simplest way I can explain it. It's like having a coworker (hence the name) you hand a task to, not a chatbot you ask a question like regular ChatGPT or Claude.

Instead of prompting back and forth, you give it a workflow and it executes in phases, step by step. The outputs are consistent, cross-referenced, and formatted the same way every single time. And one of the best parts, they can be easily reused.

For me, I created a simple prompt that runs three phases automatically: extract everything from a meeting transcript, compare it to the last call on this topic or client build (you get the idea) to surface what changed, then draft a follow-up email ready to send.

I also put in my prompt to have a NEXT RUN section at the end, basically the agent telling me exactly what to paste next week so it can run the same comparison again. Same structure, just new information layered in every time.

I don't read transcripts much anymore or paste them into ChatGPT.

🧠 How It Works

1. Fathom records the call → transcript auto-saves and I drop it into a project folder (Client A, Sponsors, Internal Builds, etc.)

2. Cowork runs the workflow → I tell if what folder to pull from (saved on Google Drive or Desktop) + the prior one and kick off the agent

3. Phase 1: Extract → it pulls every decision, action item, owner, due date, and open question from the raw transcript

4. Phase 2: Delta → it compares this call to the last one and flags what changed, what drifted, and what new risks appeared

5. Phase 3: Draft comms → it writes a copy/paste follow-up document referencing the exact action items and asks

   1. You can make the agent auto-email it to you or the team as well

6. NEXT RUN → it outputs a checklist of what to paste next week so it can update the delta automatically

This is a lightweight starting point, but Cowork goes much deeper.

If you have a Claude subscription, you already have access. All you need to do is just download the desktop app. Cowork is still in research preview, but it's already one of the most powerful agentic platforms I've used. Think Claude Code for non-technical people. Keep it simple like I did, or give it skill.md, Claude.md files and tools to make it even more powerful. I cover exactly how inside our community.

And if you want to go deep, I have a Cowork course is dropping in the next week. If you're interested in learning how to use agentic tools like this yourself, now's the time to join us in the Building AI Agents Community!

Till next week,
✌️ AP