- AI Fire
- Posts
- 🚀 Google’s Self-Optimizing Agents
Sponsored by
Google open-sourced Harness², letting AI agents improve memory, and workflows while keeping the base model frozen. Early tests showed gains up to 11.8 points.
What's on FIRE 🔥
IN PARTNERSHIP WITH ELMO
Your customers are asking ChatGPT what to buy. Does it mention you? Elmo is the open-source AI visibility tracker, within minutes, Elmo shows you:
How often AI mentions your brand & your AI Share of Voice
Which competitors AI recommends instead
The EXACT sources AI trusts in your category
Which content and sources you should fix FIRST
Visibility across ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Copilot, Grok & more
Plans from $29/mo, or self-host the entire platform for FREE.
AI INSIGHTS
Google Research has open-sourced Harness², a framework that lets AI agents improve how they work without changing the underlying model weights:
It can rewrite 8 parts of an agent harness: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins.
It does this without needing a ground-truth answer during the improvement process. Successful configurations and lessons are stored in a persistent library, so later tasks can benefit from what previous runs learned.
Developers can run improvements in parallel or sequentially across multiple refinement rounds.
Early tests showed gains of up to:
+11.5 points on WorkBuddy-Bench
+4.5 points on JobBench
+11.8 percentage points on LAB’s partial-pass rate
The experiments included Gemini 3.5 Flash and GLM-5, and performance generally improved further as the recursion budget increased.
Google released the project under Apache 2.0, with support for both OpenCode and Codex executors. It is a Google Research project rather than an officially supported Google product.
PRESENTED BY ATLASSIAN
State of Product reveals what AI still hasn’t solved
80% of product professionals say AI helps them ship faster, but customers aren’t seeing value any sooner. Atlassian’s State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasn’t kept pace to gut instinct overriding customer evidence.
RECOMMENDED GUIDE YOU’VE MISSED
There are 5 free AI tools that can make your Claude workflow much more powerful.
I've tested them with real examples, and I'll show you exactly what each tool does, when to use it, and how to connect it to your workflow:
Lavish turns Claude's plan dumps into an interactive HTML page you can annotate
Ponytail pushes Claude toward simpler solutions; now has 5 sub-skills
RTK for real 80%+ reductions on CLI output, just don't expect that on your whole bill
Caveman cuts output text by 65% in tests; real API cost savings are more like 15–25%
Small Business has 43 workflows, 27 integrations
For each tool, you'll get:
What it actually does and why you should care.
A real-world example you can try yourself.
A step-by-step tutorial to get it working.
You don't need to pay for another expensive AI subscription or rebuild your entire setup. Let's get these tools working with Claude.
TODAY IN AI
AI HIGHLIGHTS
🏗️ A new Claude Code skill turns machine ideas into surprisingly detailed digital prototypes, right down to individual bolts (link). No surprise it’s already gone viral.
🕷️ You promote one article on Instagram… and Meta could potentially scrape the whole thing from your website. Read the fine print. Now, it has been removed!?
💻 Replit says OpenAI models now outperform Claude on vibe-coded apps in its internal benchmarks. Claude models are still available, but GPT is now the default.
🔬 One long-held AI assumption might be wrong. In an informal research essay, Nick Cammarata argues models are far easier to interpret than researchers assumed.
🏰 Disney agreed to hire most of Character(.)AI’s workforce after the startup filed for bankruptcy. At least 60 employees received offers, and most reportedly said yes.
🚨 China could restrict exports across 23 categories of AI technology. The proposal adds another layer to the global AI race, while keeping publicly available tech exempt.
💰 Big AI Commitment: Anthropic is putting 5% of its shares into a new public benefit fund, with more added as its value grows. This could support nonprofits, direct aid, and AI job-impact research.
NEW EMPOWERED AI TOOLS
📧 Inbox Zero brings all your Gmail and Outlook accounts into one free inbox, with AI that sorts emails and drafts replies in your voice.
🎤 MyLens gives every person a personalized AI presentation that answers questions live, and creates slides on the spot automatically.
🔐 shim-cli protects your secrets and personal data when using AI coding agents, while letting you see exactly what they send.
💻 Tower 18 lets you jump straight to Git history, blame, branches, and more with a single terminal command, making Git workflows faster and easier.
AI BREAKTHROUGH
Elorian released Glimpse, a new visual reasoning model designed to keep looking at images and videos while it thinks.
Most multimodal models inspect an image once, convert it into tokens, then reason from that representation. Glimpse can go back to the source, zoom into specific areas, pull more video frames, and run code to verify what it sees. Elorian calls this a “visual scratch space.” Main points:
Built for images, video, charts, counting, and physical-world reasoning.
Can crop, re-check frames, compare details, and use code during reasoning.
Costs $0.60 per million input tokens and $3 per million output tokens.
In Elorian’s own tests, it beat GPT-6 Astra on 4 of 11 benchmarks, including scientific chart reasoning and short-form video captioning, while Astra still led on several harder visual reasoning tests.
Available now as a research-preview API.
Glimpse’s interesting idea is giving AI the ability to look again before answering, closer to how humans zoom in, rewind, and double-check visual evidence instead of relying on a single glance.
A CRM so smart, it updates itself.
HubSpot's CRM is so smart, it now updates itself. Calls get logged and summarized the moment they end. New leads get researched and a first outreach drafted, automatically. Deals move forward on their own, with next steps flagged before you have to ask.
Even the one repetitive task you've been meaning to fix can now run itself, no code required. Customers using it are already seeing the difference.
This isn't the CRM you have to work hard to figure out. It's the one already working before you log in.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply