
I teach the ABCs of AI Agents Tokens are going to get cheaper, use them AI team lead @squidgyai, IIT Delhi
Joined November 2013
- Tweets5K
- Following1.6K
- Followers508
- Likes155K
I broke my wrist, so I tested 9 Mac dictation apps on my real Claude Code prompts.
Same recorded audio into every app, and a Claude agent drove all of them.
Most accurate: Aqua Voice (100%)
Fastest: Parakeety and TypeWhisper (0.6 s)
Lightest: FreeFlow (117 MB)
Heaviest: Wispr Flow (1.6 GB)
I'm staying on the one that uses the least RAM, which one is your pick?
Gemini agents no longer load every tool's full schema.
Google put Tool Retrieval into preview in the Gemini API yesterday. It works with Gemini 3.6 and every Gemini 4 model.
1. Today an agent with 100 tools sends each tool's schema (its full list of inputs) on every request. That burns tokens, and the model picks the wrong tool more often.
2. Now you mark tools with defer_loading. The model sees only their names and descriptions, then loads a full schema by name when it needs it.
3. It works on remote MCP servers too (MCP is the standard way to plug a tool server into an agent). One flag defers every tool a server has.
4. No extra fee. You pay normal token rates, plus a small fixed block for the search tool.
My take: I'd keep my 5 most-used tools loaded and defer the rarely used ones. Then I'd measure, because every schema the model loads stays billed for the rest of the chat.
I post AI news here every day.
1
Google's new Gemini is built to run your agents.
Google shipped Gemini 4 Neon yesterday: $3 per million input tokens, $15 per million output, and a 1 million token context window.
1. It ranks #1 on the Artificial Analysis Coding Agent Index (58.6) and #1 on Zapier's AutomationBench (56.3%).
2. At its default (high) thinking level, a task costs 32% less than on Gemini 4 Carbon, Google's model for the hardest long projects.
3. It can act as the orchestrator. Neon plans, and faster Hydrogen subagents (helper agents that take smaller jobs) do the routine edits and test runs.
4. Cached input drops to $0.15 per million. That's a 95% cut for agents that resend the same long prompt every turn.
My take: a smart planner with fast workers is how I already build agents. I'll test Neon as the planner this week.
I post AI news here every day.
1
55
OpenAI's new GPT-6.2 has a catch for agent builders.
OpenAI shipped GPT-6.2 Astra and Sol yesterday. Astra costs $8 in and $40 out per million tokens, 20% below GPT-6 Astra.
1. Sol stays at $2/$10. It lands within 0.5 points of Astra on FrontierCode 2.0 (a test of merge-ready code) at about 1/5 the cost, both at max effort.
2. Cached input on Astra drops 60%, to $0.40. That's prompt text the API has already seen.
3. The catch: edit earlier messages mid-run and the API drops the model's saved reasoning. Quality falls and token use goes up.
4. OpenAI's own tests found both models more often stay quiet about blockers, like a missing file.
My take: before I switch, I'm making my agent history append-only and adding a check that fails loudly when a tool or file is missing.
I post AI news here every day.
2
1
53
GPT-6.2 Astra is about 20% cheaper than 6.1 Astra. I know exactly which agent gets it first: the reviewer that reads my PRs before I merge.
Last week Codex reviewed a PR with 270 passing tests before I merged it.
So Astra gets that seat for a day, and I compare what it flags against Codex.
Would you trust a cheaper reviewer?
HERE IT IS: GPT-6.2 ASTRA AND SOL
Its 20% cheaper than GPT-6.1 Astra!
OpenAI says GPT-6.2 Astra scores 5.5 points higher than Claude Fable 5.2 on Agents’ Last Exam at 5% of the cost per task.
BUT it falls short for example in Frontier Code 2.0 to Fable 5.2!
So far: Very efficient model, comparable to Fable 5.2 but looks like they are more on par than GPT-6.2 outperforming Fable!
1
1
99
30 days of Claude Code in 4 quick questions.
Anthropic already gave you a free course, Claude Code 101. I post one tip from it every day.
Day 30/30.
It's the last day, so here's a short quiz on the whole series.
1. Which mode lets Claude read your code but not edit it?
2. You finished one feature and want to start the next. Do you run /compact or /clear?
3. You need one fact from a big codebase, like which file handles login. What keeps your main context clean?
4. A PreToolUse hook (a command that runs before a tool call) needs to stop the call. Which exit code does it return?
Answer all 4 first. Then check:
Plan Mode. /clear. A subagent. Exit code 2.
Try all 4 today, then tell me which one you missed.
2
1
87
From Anthropic's free Claude Code 101 course: academy.claude.com/courses/c…
1
27
OpenAI's agent kit now asks before it acts.
OpenAI shipped Agents SDK v0.24.0-alpha.8 for Python on 16 Nov. It's a preview build, and tool approvals lead it.
1. Your app sees a tool call's inputs before it runs. A tool call is the agent using a function, like sending an email. You approve or reject it.
2. You can remember that choice for the rest of the session. No more approving the same safe call again and again.
3. Resumed runs (restarted from a saved checkpoint) keep the token budget, your cap on model usage, that they already spent. So a restart can't reset the limit.
4. It also works when you call models through your own code, like LiteLLM or a custom adapter.
My take: I'd gate every tool that writes or deletes and let read-only tools run free. Back up saved state first. Alpha.7 can't read the new budgeted checkpoints.
I post AI news here every day.
1
27
Source: github.com/openai/openai-age…
2
Five needs, five homes in Claude Code. Pick by need.
Anthropic already gave you a free course, Claude Code 101. I post one tip from it every day.
Day 29/30.
Claude Code gives you 5 places to put what you want. Each one fits a different need.
1. Rules for every session go in CLAUDE.md, a file Claude reads at the start.
2. A how-to you need only for one kind of task goes in a skill. It loads only when your request matches. A step that must happen every time, like formatting after each edit, goes in a hook.
3. Data outside your code, like your issue tracker, needs an MCP server. MCP (Model Context Protocol) is an open standard that connects Claude Code to outside tools and data.
4. A side question where you only want the answer goes to a subagent. It works in its own context and sends back a short summary.
My test: if I typed it to Claude twice this week, it gets a home today.
3
75
AWS just made MCP setup a one-key job.
AWS DevOps Agent, AWS's agent that digs into production incidents, added partner-managed MCP connectors on 16 Nov.
1. MCP (Model Context Protocol) is the standard plug that lets an agent use outside tools. You used to look up the server address, the login URLs and the permission scopes yourself.
2. Now the vendor publishes a connector in AWS's open GitHub repo. You pick an endpoint and paste only your own credentials.
3. A connector can bring the vendor's recommended skills (playbooks the agent loads) and an alert webhook (a URL their product calls when something breaks). Their alerts then start investigations.
4. Watch two things. Requests can leave your AWS Region. And removing a connector can delete the skills it installed.
My take: this is how MCP should ship. The vendor owns the config, I bring my own credentials. I'd still switch on read-only tools first.
I post AI news here every day.
3
1
55
Claude cut my 11,864 saved keyboard words to 3,584.
I had Claude clean up every word my old keyboard learned and turn it into a file my new keyboard can import. You could do the same in an evening.
The problem: I switched phones and moved to Gboard. My old keyboard had saved 11,864 words from my typing. Most of them were words Gboard already knows.
The idea was simple. Gboard already knows normal English. So I only needed the words it would never guess on its own.
Claude sorted the list for me. It dropped 6,456 common English words. It merged 589 repeats that only differed in capital letters. Then it threw out numbers, emoji, links, emails and typos.
What's left is 3,584 words. About 2,450 are Hinglish (Hindi typed in English letters) and slang. About 1,130 are names, places and brands. And before writing the file, Claude checked the format against a real dictionary file from the new keyboard. No guessing.
How to start:
1. Export the saved words from your keyboard app, if it lets you.
2. Give the file to Claude and ask it to keep only words a new keyboard wouldn't already know.
3. Ask Claude for the import format your new keyboard uses, and check it against a sample file first.
1
39
Your agent's tool list may be wrecking your cache.
Vercel shipped fx v0.0.17, its terminal coding agent, on Nov 14. In tests across 9 model families, prompt cache reads rose from 61.3% to 82.0% of later input.
1. A prompt cache lets the provider reuse the start of a prompt it already processed. Cached tokens cost less and come back faster.
2. The trick: one "toolbox" tool replaces 3 older tools. Every request now sends the same tool list, so the cache stops breaking.
3. Built-in tool definitions got 48% smaller. A new session's first request is 32% smaller.
4. Subagents work alongside the main agent and report back when done. Each gets its own model and permissions, and they nest 2 levels deep.
My take: I'm auditing my own agents for anything that changes the prompt start between turns. A stable tool list is the cheapest speedup I know.
I post AI news here every day.
2
55
Source: github.com/vercel-labs/fx/re…
10
Every Shopify store just became readable by AI agents.
Sierra released Personal Agent Protocol 1.0 (Poppy) on Nov 13. It's an open standard for how personal agents deal with companies, and it has 100+ design partners.
1. A personal agent is an AI that acts for you. Think booking a flight or returning an order.
2. Poppy is built mostly on existing standards: MCP (a way to give agents tools), OpenAPI (a common format for describing APIs) and UCP (a shopping checkout standard).
3. You sign in with OAuth and pick what your agent may do. You can change that later.
4. Shopify turned it on for every merchant. Each store now publishes a Poppy file automatically. Meta's Muse agent checks for Poppy first on every site it visits.
My take: I've written too many fragile browser scripts for agents. A file the site publishes on purpose beats my agent guessing which button to click.
I post AI news here every day.
6
3
125
Source: sierra.ai/blog/releasing-pop…
1
19


