For complex agents, enforcing behaviors like persistence, risk assessment, and proactive planning drives better reliability in multi-step workflows.
Resources 🔗
SI Template:ai.google.dev/gemini-api/doc…w
Docs:ai.google.dev/gemini-api/doc…U
AI closes the gap on doing.
Taste closes the gap on deciding.
And deciding is where all the value is created.
AI will take over the typing.
It won’t take over the knowing.
"These models are a very different kind of intelligence than human intelligence.
Human intelligence is interesting because I think that I understand things because I can introspect my own thought process to some extent and then I believe that my thought process probably works similar to other people's.
But these models are sort of like this alien form of intelligence where they can do really interesting things... but whatever internal self-reflection that they have if it exists at all is totally different from what we do."
Justin breaks down the stark differences between human intelligence and model intelligence.
@jcjohnss
From years as a Stanford professor and building ImageNet, the project that catalyzed the deep learning revolution, Fei-Fei Li joined forces with her former student Justin Johnson and a select team to start World Labs, staking their careers on spatial intelligence as the next frontier in AI.
We sat down with Fei-Fei and Justin to dig into what the “AlexNet moment” for world models looks like, why transformers are really set models rather than sequence models, how Gaussian splats, physics priors, and new hardware primitives could redefine the stack beyond GPUs, the shifting role of academia in an era of frontier labs, how grounding AI in physical reality could change robotics, medicine, and creative design forever, and their bet that spatially grounded world models will be the bridge from today’s text-only LLMs to agents that actually live and act in 3D space.
@drfeifei @jcjohnss @swyx @FanaHOVA
𝐏𝐫𝐨𝐦𝐩𝐭 𝐞𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 𝐠𝐞𝐭𝐬 𝐚𝐥𝐥 𝐭𝐡𝐞 𝐚𝐭𝐭𝐞𝐧𝐭𝐢𝐨𝐧.
But 𝐜𝐨𝐧𝐭𝐞𝐱𝐭 𝐞𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 is what separates working demos from production systems.
𝐇𝐞𝐫𝐞 𝐢𝐬 𝐡𝐨𝐰 𝐭𝐡𝐞𝐲 𝐝𝐢𝐟𝐟𝐞𝐫 𝐚𝐧𝐝 𝐰𝐡𝐲 𝐛𝐨𝐭𝐡 𝐦𝐚𝐭𝐭𝐞𝐫:
𝟏. 𝐏𝐫𝐨𝐦𝐩𝐭 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 = 𝐓𝐡𝐞 𝐐𝐮𝐞𝐬𝐭𝐢𝐨𝐧.
🔷You write clear instructions for a specific task.
🔷You define the desired format and constraints.
🔷You optimize the wording to get better outputs.
🔷This is what most tutorials teach.
𝟐. 𝐖𝐡𝐚𝐭 𝐢𝐭 𝐬𝐨𝐥𝐯𝐞𝐬.
🔷Clarifies the immediate goal.
🔷Sets output boundaries.
🔷Improves response quality through better phrasing.
🔷Works great for one off requests.
𝟑. 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 = 𝐓𝐡𝐞 𝐅𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧.
🔷You architect what information flows into the model.
🔷You design how knowledge is retrieved and structured.
🔷You manage state across multiple interactions.
🔷This is what makes AI systems actually useful.
𝟒. 𝐖𝐡𝐚𝐭 𝐢𝐭 𝐢𝐧𝐜𝐥𝐮𝐝𝐞𝐬.
🔷Dynamic retrieval from knowledge bases.
🔷Relevant examples selected at runtime.
🔷Previous conversation history.
🔷External data from tools and APIs.
𝟓. 𝐓𝐡𝐞 𝐫𝐞𝐚𝐥 𝐢𝐧𝐬𝐢𝐠𝐡𝐭.
🔷A perfect prompt with no context is like asking an expert with amnesia.
🔷Context engineering builds memory.
🔷Prompt engineering asks the right question.
🔷You need both, but context scales better.
𝐀𝐫𝐞 𝐲𝐨𝐮 𝐬𝐩𝐞𝐧𝐝𝐢𝐧𝐠 𝐦𝐨𝐫𝐞 𝐭𝐢𝐦𝐞 𝐨𝐧 𝐲𝐨𝐮𝐫 𝐩𝐫𝐨𝐦𝐩𝐭𝐬 𝐨𝐫 𝐲𝐨𝐮𝐫 𝐜𝐨𝐧𝐭𝐞𝐱𝐭 𝐚𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞?
———
👉 Follow @milvusio, for everything related to unstructured data!
𝐁𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐚𝐧 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭 𝐢𝐬 𝐞𝐚𝐬𝐲. 𝐌𝐚𝐤𝐢𝐧𝐠 𝐢𝐭 𝐩𝐫𝐨𝐝𝐮𝐜𝐭𝐢𝐨𝐧-𝐫𝐞𝐚𝐝𝐲? 𝐓𝐡𝐚𝐭'𝐬 𝐰𝐡𝐞𝐫𝐞 𝐦𝐨𝐬𝐭 𝐩𝐫𝐨𝐣𝐞𝐜𝐭𝐬 𝐟𝐚𝐢𝐥.
Here's the 7-stage roadmap to production-ready AI agents:
𝟏. 𝐃𝐢𝐬𝐜𝐨𝐯𝐞𝐫:
Start by defining the problem clearly.
Who are your users? What matters most—speed, accuracy, cost?
Identify risks like data privacy, bias, and compliance early.
𝟐. 𝐃𝐞𝐬𝐢𝐠𝐧:
Map out your agent's behavior and decision-making flow.
Will you use ReAct, Plan-and-Execute, or another pattern?
Decide between single-agent, multi-agent, or hybrid architectures.
𝟑. 𝐂𝐨𝐧𝐧𝐞𝐜𝐭:
Integrate the tools and APIs your agent needs.
Add memory systems for context.
Use frameworks like LangChain for orchestration.
𝟒. 𝐏𝐫𝐨𝐦𝐩𝐭:
Craft clear prompts with instructions and examples.
Use few-shot learning and chain-of-thought(CoT) reasoning to improve quality and consistency.
𝟓. 𝐆𝐫𝐨𝐮𝐧𝐝:
Connect your agent to reliable data sources.
Implement RAG to reduce hallucinations.
Use 𝐯𝐞𝐜𝐭𝐨𝐫 𝐝𝐚𝐭𝐚𝐛𝐚𝐬𝐞𝐬 𝐥𝐢𝐤𝐞 𝐌𝐢𝐥𝐯𝐮𝐬 to store and retrieve embeddings at scale efficiently.
𝟔. 𝐓𝐞𝐬𝐭:
Run comprehensive tests with real scenarios.
Include edge cases, adversarial inputs, and load testing.
Iterate based on failures and user feedback until performance is consistent.
𝟕. 𝐃𝐞𝐩𝐥𝐨𝐲:
Take your agent live with proper infrastructure.
Use FastAPI, Docker, or serverless platforms.
Monitor with tools like LangSmith, track costs, latency, and errors in real-time.
Building production-ready AI agents isn't just about getting code to work—it's about designing for scale, reliability, and continuous improvement.
What stage are you currently working on?
———
👉 Follow @milvusio, for everything related to unstructured data!
🚀 Launching DeepSeek-V3.2 & DeepSeek-V3.2-Speciale — Reasoning-first models built for agents!
🔹 DeepSeek-V3.2: Official successor to V3.2-Exp. Now live on App, Web & API.
🔹 DeepSeek-V3.2-Speciale: Pushing the boundaries of reasoning capabilities. API-only for now.
📄 Tech repohuggingface.co/deepseek-ai/De…NuG0
1/n
Anthropic is acquiring @bunjavascript to further accelerate Claude Code’s growth.
We're delighted that Bun—which has dramatically improved the JavaScript and TypeScript developer experience—is joining us to make Claude Code even better.
Read more: anthropic.com/news/anthropic…
Banger paper for agent builders.
Multi-agent systems often underdeliver. The problem isn't how the agents themselves are built. It's how they're organized.
They are mostly built with fixed chains, trees, and graphs that can't adapt as tasks evolve.
But what if the system could learn its own coordination patterns?
This new research introduces Puppeteer, a framework that learns to orchestrate agents dynamically rather than relying on handcrafted topologies.
Instead of pre-defining collaboration structures, an orchestrator selects which agent speaks next based on the evolving conversation state. The policy is trained with REINFORCE, optimizing directly for task success.
Rather than searching over complex graph topologies, they serialize everything into sequential agent selections. This reframing sidesteps combinatorial complexity.
What emerges is surprising: compact cyclic patterns develop naturally. Not sprawling graphs, but tight loops where 2-3 agents handle most of the work.
The remarkable part is that the system discovers efficiency on its own.
Results:
- On GSM-Hard math problems: 70% accuracy (up from 13.5% for the base model alone).
- On MMLU-Pro: 83% (vs 76% baseline).
- On SRDD software development: 76.4% (vs 60.6% baseline).
These gains come with reduced token consumption. The paper shows that token costs consistently decrease throughout training while performance improves.
They also prove the agent selection process satisfies Markov properties, meaning the current state alone determines the optimal next agent. No need to track full history.
Why it matters for AI devs: learned simplicity beats engineered complexity. A trained router with a handful of specialized agents can outperform elaborate handcrafted workflows while cutting computational overhead.
With state-of-the-art reasoning, richer visuals, and deeper interactivity, Gemini 3 is more intuitive, more powerful, and more personalized.
Start exploring at gemini.google or in the app by selecting “Thinking” from the model picker.
