MATE: The Open-Source Orchestration Layer Google ADK Left Out
A developer shipped MATE (Multi-Agent Tree Engine) for Google ADK this weekend — open-source, with a built-in dashboard, MCP support, persistent memory, and compatibility with 50+ LLM providers. Google ADK gives you the raw agent primitives; MATE gives you the production scaffold around them. It targets exactly the orchestration gap teams hit when they move beyond single-agent demos.
Why it matters: Start with MATE's architecture before scoping your next multi-agent build — it solves the orchestration problems you'll otherwise spend weeks hitting on your own.
Read more →
Simon Willison on Lenny's Podcast: What Production Agents Actually Look Like
Simon Willison appeared on Lenny's Podcast — 500K+ subscribers — to break down how agentic systems work in production: context management, tool use, error recovery loops, and the architectural decisions that separate deployed products from demos. The episode published April 2 and is already circulating among product leaders who will be bringing its vocabulary into meetings by end of week.
Why it matters: Get fluent in the patterns now — the people who fund and commission your work are about to start asking about them.
Read more →
Gemma 4 + LM Studio + Claude Code: A Fully Local Coding Agent That Costs Nothing to Run
George Liu published a hands-on guide to running Google's Gemma 4 locally via LM Studio's headless CLI, then wiring it into Claude Code as an inference backend. The setup delivers a frontier-class coding agent entirely on your own hardware — no API calls, no data leaving your machine, no per-token cost. The guide is practical and reproducible.
Why it matters: If you've been waiting for local models to clear the bar for real code work, this is the benchmark worth running this week.
Read more →
|
Pattern Watch
This weekend's shipping spree shows agentic engineering moving from theory to production-ready tooling. MATE provides the missing orchestration layer, while local setups like Gemma 4 eliminate cost barriers — together they enable teams to build real multi-agent systems without waiting for platform vendors.
|
Radar
|
Agent error memory
Thumbs-up/thumbs-down feedback that stops agents from repeating the same mistakes — simple UX, real reliability gain.
Link →
|
|
94-skill Windows agent desktop
NousResearch's Hermes wrapped in a full WinUI 3 app with multi-agent profiles and a soul system — second side project, ships working software.
Link →
|
|
Day 34: AI trading lab with $500 real money
A silent bug froze all progress for weeks; the post-mortem is the clearest account yet of what real money plus real agents looks like when things go wrong.
Link →
|
|
1,000-stakeholder simulation engine
AI sim that stress-tests business decisions against 1,000 synthetic stakeholders before you commit — novel approach to de-risking strategy with agents.
Link →
|
|
Codex goes token-based
OpenAI shifting Codex pricing from per-message to API token usage — recalculate your agent cost model if Codex is in your stack.
Link →
|
|
Tool of the Day
Goose by Block
Open-source coding agent from Block — Jack Dorsey's company, formerly Square — that runs locally with modular extensions for GitHub, filesystem access, and custom tools. Trending on GitHub this week, which means the community is actively stress-testing it. Block engineers production-grade software; this isn't a weekend experiment.
GitHub →
|
|
Under the Hood
Today's edition: 173 sources scanned by Atlas (DeepSeek) → Curator (Claude) selected the stories → Scribe (Claude) wrote the draft → Mercury (DeepSeek) formatted for delivery. Atlas: $0.003425 | Claude agents: ~$0 (Max subscription). The RSS feeds returned high relevance scores on older content — today's top picks came from fresh Reddit SideProject posts, with Willison's Lenny's appearance as the only RSS story that cleared the freshness bar.
|
|
