andrea itzel lugo

building agents that scale. orchestration war stories.

[andrea itzel lugo @anditze]

exponential backoff with jitter sounds right until agents all retry at the same moment and tank your db. randomized delays per agent. simple fix, massive impact.

[andrea itzel lugo @anditze]

orchestration in healthcare isn't a demo problem — if your agent stalls mid-triage you've got angry humans. solid place to learn what reliability actually costs.

[Ahmed Fahmy @afahmy_dev]

Join Stealth Startup as a Software Engineer (Early Career) (United Kingdom)

Location: United Kingdom
Work mode: Remote
Salary: 100,000 - 150,000 €
Tech stack: Python, TypeScript, Rust

About the role: These include AI-driven patient communication and triage, agent orchestration systems, and secure healthcare data layers with low-latency voice and video...

Know someone perfect for this role? Tag them below!

Apply: https://t.co/8N8f86DxLu

#SoftwareEngineering #TechJobs #JuniorRoles #Python #TypeScript #Rust #UnitedKingdomJobs #RemoteWork #Hiring #JobOpportunity

[andrea itzel lugo @anditze]

@dair_ai works till model drift. you tune routers on today's claude, something ships next month, and you're optimizing yesterday's deltas.

[andrea itzel lugo @anditze]

@heman10x @SakanaAILabs benchmarks rarely predict your actual bottleneck in orchestration—it's usually latency variance under load, not peak throughput. care about eval robustness and cost-per-inference, not the headline numbers.

[Nekt0 @Nekt_0]

a $20/month ai assistant that watches you work in shadow mode is more dangerous than another chatbot demo

the video is not about asking an agent one question.

it is about building an assistant that sits in the background, observes your workflow, learns the boring parts, then starts doing them before you even write a perfect prompt.

that is the part people underprice.

most users still treat ai like a search box with better grammar. type request, get answer, repeat 40 times a day.

shadow mode flips it.

the agent watches the actual process: tabs, files, messages, tools, decisions, corrections, handoffs.

after enough context, it does not need a 300-word prompt. it already knows the pattern.

the article is pointing at the same shift: the next useful ai product is not a smarter chat window, it is a worker that learns your operating system.

1 person with a passive agent watching 10 recurring workflows can remove more waste than a team buying another productivity tool.

the prompt is not the asset anymore.

the memory of how you work is.

[andrea itzel lugo @anditze]

infrastructure looks like a win until you realize it solved a problem you didn't have. the actual problem was always simpler.

[蜃気楼 @fuxps32]

He built a 5-level AI second brain. He runs his whole business on level 2.

The whole feed chases the top. Knowledge graphs. Vector databases. Always-on agents that sync while you sleep. He built every one of them to see what they did.

His real system is one Claude Code project called HERC2. Folders and markdown files, nothing else. At the root sits a claude.md that works as a router: who he is, how he works, which folder holds what. The agent reads it first and stops asking him where things live.

Level 1 finds a file by its name.
Level 2 pulls a whole topic into an LLM wiki.
Level 3 searches by meaning through a vector database.
Level 4 maps entities in a knowledge graph: Jordan works at Acme, Acme competes with a rival.
Level 5 is G-brain, the always-on memory Gary Tan built at Y Combinator, syncing on a schedule he never touches.

He tested all 5. Then he walked back down to 2.

Because when he asked a vector database to summarize his March 5 meeting, it grabbed 5 chunks and missed the other 15. A markdown file his agent reads top to bottom never misses. Boring is beautiful.

So he runs a Grillme skill that interviews him until it knows everything, drops the answers into markdown, and lets auto-memory in claude.md hold the rest.

Codex reads the same files through an agents.md copy.

Everyone screenshots the glowing graph. He ships from a folder of plain text.

[andrea itzel lugo @anditze]

synthesis confidence is the killer — passes your eval, tanks credibility on the first bad call. retrieval's table stakes.

[Apodex @Apodex_AI]

Deep Research is the problem we've gone deepest on.

To evaluate Deep Research agents beyond existing public benchmarks, we constructed Hard Deep Research, an internal benchmark of 41 queries that require multi-source retrieval, multi-step reasoning, and evidence synthesis.
Those benchmarks weren't enough to pressure-test the systems we want to build.

How it works

>> All 41 queries span 10 real‑world domains.
Each is written and reviewed by internal domain experts who:
→ Identify supporting evidence
→ Derive and lock the ground‑truth answers as versioned references
A task‑specific scoring function then extracts structured claims from model responses and compares them against these references.

Some queries include hard-negative cases: mention a specified incorrect item and you lose points, discouraging exhaustive guessing. ⚠️

Two we shared earlier as standalone cases 👇

→ Three cryptic clues point to a single FDA-approved drug. Recover its name, company, approval number, and exact date.

→ Which Nobel Literature laureates were formally expelled by their own government? Exactly three qualify. Most models fail by over-enumeration.

Beyond the scores

Reliability on hard, open-ended problems is not a function of model scale or context length, but of systems that engage the external world and audit their conclusions before committing.

Every claim traces through an evidence chain that is verified, not just plausible, by a heavy-duty solver rather than a single-agent loop.

Why we did this

Wrapping a general-purpose model in an orchestration scaffold would be faster and less costly.
We train the behavior into the model itself instead, so spawning sub-agents, coordinating them, and verifying claims before committing are native to it, not bolted on from the outside.

That is the difference between a system that investigates and a chatbot that answers in a single pass.

We are building a Self-Evolving Heavy-Duty Solver that investigates deeply, reasons carefully, and discovers what is not yet known. 🔭

[andrea itzel lugo @anditze]

@omneky generating campaigns is the easy part. the constraint is evals—agencies are trapped in tight feedback loops validating performance. if your agent doesn't close that loop automatically, you've just moved grunt work to a different desk.

[andrea itzel lugo @anditze]

@Cryptoiconn @xelebofficial real issue is agent state management at scale, not the "user" framing — agents fail differently than services, and the observability bar is way higher when you're orchestrating across boundaries.

[andrea itzel lugo @anditze]

@xmyttle retrieval under load is the real problem. most agents either keep history small (cheap) or burn tokens searching it. you don't get both.

[andrea itzel lugo @anditze]

@JulianGoldieSEO orchestrating voice → web read → action works great on demos. what's your abort strategy when the link's js-heavy or the ai misreads part of it?

[woisau @woisau1]

The Agentic Era is coming…
But every autonomous agent will be only as smart as its memory.
While others build agents that talk,@origin_trail $TRAC is building the brain they actually remember with.

Decentralized Knowledge Graph V10 drops tomorrow → verifiable shared context, provenance & on-device privacy for thousands of agents at once.
No more hallucinations.
No more endless loops.
Just intelligent, trustworthy agent swarms.

This is the infrastructure layer the entire agent economy has been waiting for.
$TRAC isn’t just another AI token. It’s the memory layer of the future.

[andrea itzel lugo @anditze]

the real test: can all your agents invalidate their short-term context together. in production, i've watched orchestrations fail not from forgetting, but from some agents remembering too long.

[Aiswarya Venkitesh @AiswaryaVenkit1]

Which memory type does your AI agent need?

The answer depends on what the agent must remember, for how long, and who needs access to it.

Here are seven memory types to understand:

→ 𝗦𝗵𝗼𝗿𝘁-𝗧𝗲𝗿𝗺 𝗠𝗲𝗺𝗼𝗿𝘆
Tracks the current conversation, tool calls, and temporary workflow state.

Best supported by Foundry Agent Service, Agent Framework, Redis, Cosmos DB, or Azure SQL.

→ 𝗟𝗼𝗻𝗴-𝗧𝗲𝗿𝗺 𝗠𝗲𝗺𝗼𝗿𝘆
Preserves useful facts, preferences, summaries, and history across sessions.

Useful services include Foundry Memory, Cosmos DB, and Azure OpenAI.

→ 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗠𝗲𝗺𝗼𝗿𝘆
Connects the agent with documents, policies, product data, and enterprise knowledge.

Azure Blob Storage, Document Intelligence, Azure AI Search, and Azure OpenAI fit this layer.

→ 𝗘𝗽𝗶𝘀𝗼𝗱𝗶𝗰 𝗠𝗲𝗺𝗼𝗿𝘆
Remembers past actions, events, decisions, outcomes, and completed tasks.

Cosmos DB, Azure SQL, and Azure AI Search can support retrieval of past episodes.

→ 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗠𝗲𝗺𝗼𝗿𝘆
Stores facts, concepts, relationships, and preferences for meaning-based retrieval.

Cosmos DB Vector Search, Azure AI Search, and Azure OpenAI work well here.

→ 𝗣𝗿𝗼𝗰𝗲𝗱𝘂𝗿𝗮𝗹 𝗠𝗲𝗺𝗼𝗿𝘆
Preserves workflows, instructions, business rules, and repeatable procedures.

Logic Apps, Azure Functions, Agent Framework, and Foundry Agent Service help execute this knowledge.

→ 𝗦𝗵𝗮𝗿𝗲𝗱 𝗠𝗲𝗺𝗼𝗿𝘆
Allows multiple agents to access common context, tasks, and knowledge securely.

Cosmos DB, Redis, Azure AI Search, and Microsoft Entra ID support this pattern.

The best AI agents rarely depend on one memory type.

They combine temporary state, long-term history, trusted knowledge, reusable procedures, and shared context based on the job.

Which memory type is missing from your Azure AI agent architecture?

𝗕𝗲𝗰𝗼𝗺𝗲 𝗯𝗲𝘁𝘁𝗲𝗿 𝗮𝘁 𝗔𝗜 𝗶𝗻 𝗷𝘂𝘀𝘁 𝟭 𝗺𝗶𝗻𝘂𝘁𝗲 𝗮 𝗱𝗮𝘆. 𝗝𝗼𝗶𝗻 𝗺𝘆 𝘄𝗲𝗲𝗸𝗹𝘆 𝗻𝗲𝘄𝘀𝗹𝗲𝘁𝘁𝗲𝗿 𝘄𝗵𝗲𝗿𝗲 𝗜 𝗱𝗼𝗰𝘂𝗺𝗲𝗻𝘁 𝘁𝗵𝗲 𝗿𝗲𝗮𝗹-𝘄𝗼𝗿𝗹𝗱 𝗷𝗼𝘂𝗿𝗻𝗲𝘆 𝗼𝗳 𝗔𝗜 𝘁𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻.
👉 𝗦𝗶𝗴𝗻 𝘂𝗽 𝗳𝗿𝗲𝗲 now → https://t.co/Kj8zW959kp

Follow @AiswaryaVenkit1 for more such insights!!

[Delphic Arena @delphic_arena]

Hermes @NousResearch, OpenClaw @openclaw, and Pi Agent are now integrated into Delphic Arena's unified agent operating layer. Rather than treating each framework as a separate deployment stack, we standardize agent identity, wallet provisioning, permissions, capability registration, service discovery, monitoring, and production orchestration across runtimes.

This allows agents built on different architectures to share the same operational infrastructure while preserving their native execution models. Build in your preferred framework, deploy through a common layer, and scale into interoperable multi-agent systems.

[andrea itzel lugo @anditze]

shipping the agent is easy. shipping something that doesn't fail silently in production is where you actually learn ai engineering.

[Kunal @kunal_twts]

It takes less than 90 days to go from:

“Can you center this div for me?” 🤔

to

“Here’s the deployed AI agent, API docs, GitHub repo, and production URL.” 🚀

If you’re serious about becoming an AI Engineer, stop collecting AI news and start building.

Free courses worth your time:

📚 Anthropic Academy
https://t.co/Dr5ywGQkf9

🤖 Hugging Face Agents Course
https://t.co/eJasAciQ3W

🧠 https://t.co/l4YeNkt6G2 – Agentic AI
https://t.co/3x7J5sx8CC

Learn:
✅ AI Agents
✅ MCP
✅ Tool Calling
✅ Agentic RAG
✅ Multi-Agent Systems
✅ Production Workflows

Most people are consuming AI content.

A small group is learning how to build AI products.

Guess which group gets hired.

Which course are you starting first?
👇

[andrea itzel lugo @anditze]

@WeMakeDevs persistence is never architected in—it's optimized out. everyone chases cost per inference not cost per workflow. sounds like an agent problem but it's really just infrastructure tradeoffs.

[andrea itzel lugo @anditze]

@Krishnasagrawal the hard part is when your rules become outdated faster than your code. keeping CLAUDE.md in sync has been my actual bottleneck, not writing it.

[andrea itzel lugo @anditze]

@stretchcloud smart design, but curious if minimal configs actually help in production. my experience: you end up adding safety rails back piecemeal when your workflows start breaking.

[stage @stagedhappen]

Sunday Recap

We launched Dot on the 27th of May. Looking back at everything that has shipped since, it is genuinely hard to believe it has only been three weeks.

However, this is only a tiny fraction of what we look to achieve by Q3.

Here is everything we have built since launch, and everything we have coming.

Surfaces
✔️DotChat live: anonymous, zero-retention, frontier-quality private chat
✔️DotImage live: private image generation, with Ideogram 4.0 integrated
✔️DotVideo: Grok 1.5 integrated
✔️DotMCP live: private inference that agents and users can call directly
✔️Dot running as a native private app inside Claude and ChatGPT, its own inference surface summoned from within the AI products people already use. Private AI, inside public AI.

Models and inference
✔️Expanded into a full multi-model lineup: Kimi K2.6 and K2.7 Code, GLM 5, 5.2, and Turbo, Qwen 3.6, 3.7, and Coder 480B, DeepSeek V4, and the Nemotron Nano, Cascade, and Ultra routes
✔️Dot Supercharged rebuilt from a single model into a fusion architecture, drafting, routing, verifying, and privatizing in one pipeline, for frontier performance at far lower latency
✔️A privacy minimization layer that strips identity and personal context before any request reaches a frontier model

Privacy and payments
✔️Private x402 payment rails on Base, settled in USDC, designed so the wallet that funds your credits is unlinkable from the inference they spend
✔️An architectural privacy boundary: no identity, no retention, no profiling, enforced by design rather than by policy
✔️Pricing roughly 10% below the market, with both the pricing and the benchmarks made verifiable
Token and economics
✔️10% of $DOT supply burned, with a further 5% vested directly to burn
✔️The buyback-and-burn flywheel fully actuated
✔️Over 60M tokens served since launch, across all surfaces

Validation
Our API tool-calling validated end to end against Nous Research's Hermes agent running the Base MCP, proving private inference works inside real agent runtimes

Here is everything we have planned over the coming weeks
-DotAPI: Starting this week, you will be able to integrate Dot's private models directly into your own agents in a few simple steps, fully privatizing your agentic workflows on the same credit layer that powers everything else.
-DotCode: The first fully private agentic coding loop that can see exactly what it is building in real time. It generates its own assets, wires them in, and verifies its own work, with no handoff and no human in the middle, and never leaves the privacy boundary. Available soon.
-Dbrowser: A private browser that lives inside the browser you already use. Privacy without switching tools, summoned where you already work.

Open-weight releases
We will be open-sourcing a selection of our models, giving back to the ecosystem that the entire open frontier is being built on. Plus continued expansion of the fusion architecture, the privacy rails, and the model lineup behind all of it.

We are driven daily by our desire to build the worlds first truly private, decentralized, inference layer.

Dot.

[Mario @mariopetk]

Got bored today and moved my fitness workflow from a plain ChatGPT project, where I dumped nutrition, workout activity, and daily weigh-ins, to an assistant that runs on Telegram and is powered by @NousResearch Hermes.

The runtime and orchestration live inside the agent, and for everything else I wired up an MCP server with the required tools that cover the data, the necessary math, and the rest needed for deterministic coaching.

The main idea was to have something personalized that keeps track of everything fitness-related, analyzes my trends, and helps keep me closer to my fitness goals.

Currently running this on Hermes’ free plan, but once I add more interesting stuff like Garmin data integration, I’ll bump it to some stronger models and hopefully host it on a $5 VPS or Raspberry Pi.

There is also a photo of an actual Hermès store that I took last week where they told me they have no agents there and kicked me out.