- unwind ai
- Posts
- Google ADK Visual Agent Builder
Google ADK Visual Agent Builder
+ Run 100 Large Models on a single GPU, Replit AI Integrations
Shubham Saboo & Gargi Gupta
November 11, 2025
In partnership with
Today’s top AI Highlights:
& so much more!
Read time: 3 mins
AI Tutorial
Learn about agentic design patterns, tool integration with MCP and A2A, multi-agent systems, RAG, and Agent Ops.
The 5-level taxonomy:
Level 0: Core reasoning system (LM in isolation)
Level 1: Connected problem-solver (LM + tools)
Level 2: Strategic problem-solver (context engineering + planning)
Level 3: Collaborative multi-agent systems
Level 4: Self-evolving systems that create new tools and agents
The paper covers real implementation details: how to handle context engineering, build memory systems, deploy with proper observability, and scale from one agent to enterprise fleets.
Advanced sections explore agent evolution, simulation environments, and case studies like Google Co-Scientist and AlphaEvolve.
Zero fluff. 100% free. Check it out now!
Latest Developments
Drag, drop, describe, deploy.
Google’s Agent Development Kit (ADK) release comes with Visual Agent Builder, a browser-based interface that lets you visually design, configure, and test complex multi-agent systems.
No more hand-crafting YAML. No more syntax errors.
The browser-based IDE combines a drag-and-drop canvas, a configuration editor, and a Gemini-powered AI that writes your agent setup from plain English descriptions. You can prototype with LoopAgents, SequentialAgents, and ParallelAgents, assign tools like Google Search, and configure specialized sub-agents, all through conversation, test it immediately in the built-in interface, and export production-ready code.
The beauty is that everything you do in the Visual Builder generates proper ADK YAML configurations under the hood. You can export them, version control them, and deploy them just like code-based agents.
Key Highlights:
Visual Agent Mapping – The canvas displays your complete agent hierarchy as interactive nodes showing root agents, sub-agents, tools, and their relationships, updating in real-time as you make changes through either the AI assistant or configuration panels.
Zero YAML Prototyping – Build complex workflows with specialized sub-agents, each using different models (Gemini 2.5 Pro for reasoning, Flash for speed) and tools, without manually editing nested YAML structures or debugging syntax errors.
Natural Language Configuration – The AI assistant handles clarifying questions, suggests best practices, and generates complete project structures including proper instructions, tool assignments, and hierarchical relationships based on conversational input.
Callback Management Interface – Configure all six callback types (before/after agent, model, tool) through the UI, plus built-in tool integration with searchable dialogs for Google Search, code executors, and memory management.
Quick Setup – Install with
pip install --upgrade google-adk, runadk web, and access the Visual Builder atlocalhost:8000/dev-ui.
Turn AI Into Your Income Stream
The AI economy is booming, and smart entrepreneurs are already profiting. Subscribe to Mindstream and get instant access to 200+ proven strategies to monetize AI tools like ChatGPT, Midjourney, and more. From content creation to automation services, discover actionable ways to build your AI-powered income. No coding required, just practical strategies that work.
Context bloat is killing your agents.
That search from 20 steps ago? Still taking up tokens. The 8,000-character file it read once? Still there. This all actively degrades performance, making agents slower, more expensive, and less intelligent with each interaction.
CAMEL AI has been working on solutions to this problem, focusing on a straightforward context engineering principle: only feed the agent what it needs to complete its task.
They have been using three key techniques. First, context summarization that compresses bloated conversations while keeping critical information intact. Second, workflow memory that captures how agents solved problems so they don't waste time repeating the same discovery process. Third, experiments with tool output caching revealed important lessons about the balance between efficiency and accuracy.
The team has documented everything and created GitHub issues in their open-source agent framework.
Key Highlights:
Context Compression - CAMEL's summarization extracts the main user request, pending work, current progress, and critical errors while retaining a curated list of highly informative user messages. This reduces reliance on potentially unreliable LLM-generated summaries. You can trigger summarization automatically at token thresholds, manually via API, or let agents access it as a toolkit function.
Learning from Experience - Workflow memory captures problem-solving strategies including task steps, tools used, failure recovery methods, and categorization tags. When agents encounter similar tasks, they retrieve relevant workflows through simple filename matching or agent-selected filtering. No RAG is involved to maintain a small, manageable set of dynamic memory files per agent.
The Caching Cautionary Tale - CAMEL tested storing large tool outputs externally (keeping only previews and retrieval instructions in context), achieving 94% token savings. However, they reverted this feature after discovering agents would make decisions based on incomplete preview data and struggle with the cognitive load of tracking when to retrieve full outputs. This is a great example of too much optimization hurting accuracy.
How you can contribute - The team has documented their techniques, published research showing significant performance gains, and opened specific GitHub issues for prompt improvements, workflow enhancements, and caching refinements.
Jump in, challenge yourself, ship fixes, and make agents remember better. Their framework is completely open-source so the entire ecosystem benefits.
Quick Bites
The highest accuracy web search API built for AI agents
Parallel AI launched its Search API, built specifically for AI agents rather than human browsing patterns. Instead of ranking URLs for clicks, it optimizes for token-relevance, delivering the most information-dense excerpts directly into an agent's context window. On multi-hop reasoning benchmarks like BrowseComp, it hits very high accuracy with the lowest total cost and end-to-end latency, outperforming alternatives like Exa, Tavily, and even GPT-5's browsing capabilities.
Replit AI Integrations with 300+ AI models - no API keys
Replit just killed the most annoying part of building AI apps: the setup. Their new AI Integrations give you instant access to 300+ models from OpenAI, Anthropic, Gemini, Meta, and others. No API keys, no account juggling, just tell their Agent what you want and it wires everything up automatically. Usage tracking and billing are all rolled into your Replit account. Available now.
Gemini Docs MCP Server for the entire Gemini API documentation
Google has released Gemini Docs MCP Server that gives you local, searchable access to Google's entire Gemini API documentation through a lightweight STDIO server. It uses SQLite with FTS5 for fast queries. The setup is instant - it runs via uvx, so no installation is required. It ships with three focused tools: search_documentation, get_capability_page, and get_current_model. Works with Claude Code, Gemini CLI, Codex, and any MCP client.
Kimi K2 Thinking generates a full novel from one prompt
The developer community is stress-testing Kimi K2 Thinking's extended reasoning and sequential tool-calling capabilities, and the reviews are raving. Developer Pietro Schirano released Kimi-writer, an autonomous agent that writes full novels from single prompts, complete with project planning, file management, and automatic context compression when approaching the 200K token limit. His demo generated 15 interconnected sci-fi stories in one session, with the agent independently structuring chapters and managing its workspace without human intervention. The project and the demo are worth checking out!
Gamma API now available to all
Presentation tool Gamma just closed a Series B at $2.1B valuation and hit $100M ARR with 50 people. That's $2M per employee (very impressive)! And here’s an even bigger developer story - their Generate API is now available to all, letting you programmatically spin up presentations, websites, and documents from any text input, with support for custom themes and 60+ languages. Use it via direct API calls or via platforms like Make and Zapier.
Tools of the Trade
flashtensors - Run 100 large models on a single GPU with minimal impact on time to first token. It is an inference engine that loads models from SSD to GPU VRAM 4-10× faster than standard loaders, enabling sub-2-second coldstarts for model hotswapping.
cascadeflow - Reduce your AI provider costs by 30-65% with just 3 lines of code. This open-source library routes AI queries to cheaper models first, validates the response quality, and escalates to expensive flagship models only when quality fails. Available in Python, TypeScript, and n8n community node.
Tensorlake - A Document Ingestion API that converts unstructured documents (PDFs, DOCX, spreadsheets, images) into markdown or structured data using proprietary layout detection and table recognition models. It also offers a serverless workflow runtime for building data processing pipelines with Python durable functions that automatically scale and resume from checkpoints.
Awesome LLM Apps - A curated collection of LLM apps with RAG, AI Agents, multi-agent teams, MCP, voice agents, and more. The apps use models from OpenAI, Anthropic, Google, and open-source models like DeepSeek, Qwen, and Llama that you can run locally on your computer.
(Now accepting GitHub sponsorships)
Hot Takes
That’s all for today! See you tomorrow with more such AI-filled content.
Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!
PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉
Reply