Many people still don't have a proper understanding of what AI agents are

So, here is the explanation:

🔗 Watch this https://youtu.be/oP6DS_x5K0Y?si=pTzcxmaBR2yIqebC

My TL;DR about agents:

Basically, AI agent wraps LLM with a few components:

1. Tools - to act, external capabilities to perform some action other than just generating some text (eg. web search, call APIs, run code in a sandbox, send email), mostly MCP tools. Tool = brain.

2. Memory - to remember things, long term memory, because agent cannot remember everything in 1 context window (context window = short term memory)

3. *Reasoning and planning loop - to get the task done, turns LLM into agent with continuous cycle, agent breaks big task into small sub-tasks (simulating thinking system #2 from "Thinking fast and slow"). The loop of "reAct": think → act → observe → repeat.

4. Perception - the ability to see and hear environment by monitoring chat/database/website.

AGENT = LLM + Planning + Memory + Tools

Remember: the most important component is reasoning loop 🔥 which is a continuous loop of thinking and acting.

More to read:
• https://huggingface.co/learn/agents-course/
• https://huggingface.co/learn/mcp-course/
• https://anthropic.skilljar.com/introduction-to-model-context-protocol
• https://anthropic.skilljar.com/model-context-protocol-advanced-topics

Must read:
• https://modelcontextprotocol.io
• https://geminicli.com/docs/
• https://platform.claude.com/docs/en/home

How to build own agents?
• Agent development kit: https://google.github.io/adk-docs/ - try to build own agents.
• Agent SKD by claude: https://platform.claude.com/docs/en/agent-sdk/overview
• https://docs.langchain.com/

Deep dive:
• https://roadmap.sh/ai-agents
• https://roadmap.sh/prompt-engineering

@rustamzokirov

Gemini 3.1 Pro is released today

https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/

Who is hiring in Munich right now?

Best thing to do right now? Adapt.

Why?

So that AI doesn't take your job. Or, at least, to have a window (nobody knows how wide it is).

AGI is probably 1-5 years away. I used to think people like Dario were overhyping it. This thing is moving fast and it's going to change everything economically. Prepare for the worst case scenario, the shift is fast and exponential.

> 50% of entry-level white-collar jobs could be disrupted within the next 1–5 years

> https://x.com/hamptonism/status/2027457322088538189?s=20

> https://x.com/hamptonism/status/2027462406486766006?s=20

Start trying to be unreplaceable:
• you can use AI to do the work of 10 people
• take some niche, be the best there

Telegram bots can now stream responses to users in real time.

Great for AI assistants, yes.

https://t.me/durov/475

Want to learn how AI agents are built at Yandex?

The Yandex School of Data Analysis is running a free online intensive — April 6-10, 2026.

Agents Week is a series of live video sessions where Yandex engineers walk you through:
• How to design modern AI agents from scratch
• How to take them all the way to a production-ready solution

All registered participants get access to the stream. A certificate is awarded to those who pass the selection and complete the final assignment.

✅ Participation is free
📅 Registration closes April 9, grab your spot via the link — https://agentsweek.yandex.com

Disclaimer: Yandex and ANO DPO "Yandex School of Data Analysis" do not provide educational services under the "Yandex Educational Technologies" project as part of this intensive.

Sponsored content

Build agents with Agent Development Kit (adk.dev)

ADK is Google's open-source framework for building, evaluating, and deploying multi-agent systems.

Works with Gemini, Claude, any other LLM. Multi-agent by design. And can easily communicate with other agents build with Claude agent kit, LangGraph, CrewAI.

Docs: https://adk.dev

To learn:
• Introduction to Agents and Google’s Agent Ecosystem https://www.skills.google/paths/3546

• Fundamentals of Agent Development Kit (ADK) https://www.skills.google/paths/3473

• Develop Agents with Agent Development Kit (ADK) https://www.skills.google/paths/3545

Yandex Math and Coding Contest (CMC) — now open in Uzbekistan!

CMC is an olympiad-level coding and math competition by Yandex and the Yandex School of Data Analysis — a chance to challenge yourself, meet strong peers, and unlock job opportunities at Yandex.

Who can join?
University students (18+), developers, engineers, and IT professionals in Uzbekistan.

Format:
• Free to enter
• 2 stages: online round for everyone + offline finals for top participants
• Up to 10 algorithmic and math problems per stage
• Contest language: English

Prizes for winners:
Fast track to Yandex ML School Uzbekistan, an interview for an internship at Yandex Uzbekistan, and other prizes from Yandex.

👉 Learn more and apply
🚩 Apply by April 25, 2026

Disclaimer: Yandex does not provide educational services as part of this project.

Sponsored content

I'm open-sourcing pyclaudir project, AI agent harness

🔗 Check it, run it: github.com/Rustam-Z/pyclaudir

Run self-evolving AI assistant. It lives in your Telegram. Remembers every conversation. Schedules its own work. Messages you first when something matters: a Jira ticket due, a build that broke, etc...

Example, give a task: "research global tech trends, find problems worth solving in Uzbekistan, build me a prototype." All will be done, your creativity is the limitation.

To test: Join @rustamz_workshop group, Luna, my assistant is already there.

Anyone open for contributions

@rustamzokirov

What you can do with it today:

• Drop it in your team chat → it tracks who's blocked, what shipped, writes the standup
• DM it → personal chief of staff with reminders, notes, daily check-ins
• Give it shell + code access → it ships PRs while you sleep
• Point it at a topic → it watches your messages, scrapes pages, pings you when something turns
• More examples in readme, check code

It is easy to extend, you can connect any existing MCP, skill. Refer to readme.md to understand how it works. Also check docs.

Runs on your laptop or any small VPS.

Works with your existing Claude Code subscription.

Technical depth, for the human devs

Pyclaudir under the hood:
• Python harness, runs claude code as a subprocess via local MCP, so it pipes your request to CC (next step: rewrite project for open source models, don't rely on CC)
• Agents have self-reasoning loop before responding to message
• Local MCP has many tools like sending messages, scheduling tasks, taking notes, etc
• Persistent memory: sqlite + plain markdown files
• One process per bot, run a fleet with separate pyclaudir processes
• plugins.json — you can connect external MCP servers, or turn off local MCP tools and skills
• skills/ — add agent skills
• You can turn on shell access, and your assistant will be able to write code, run it, test it — basically do everything autonomously, depending on your workflow setup
• access.json — for controlling assistant access to groups and DMs
• Runs anywhere, docker image is there already

So, the project is easily extendable, that was the goal, you can take what is done for nanoclaw, openclaw, and setup for pyclaudir.

The coolest feature: add your assistant to group for collaboration with team. The assistant remembers things, learns from you, and you manage everything from Telegram.

Next stop, San Francisco, Silicon Valley ✈️🇺🇸

Prompt engineering, context engineering, harness engineering — what's the difference?

Prompt engineering — the words you write to the model. "Do this, not that. Here's an example. Answer in JSON."

Context engineering — everything the model sees on a given turn. System prompt, past messages, files, tool descriptions, memory. You decide what goes in and what stays out, too much info is bad, too little is also bad.

Harness engineering — the system around the model. The loop, the tools it can call, permissions, sandboxing, what happens when something fails, how sub-agents talk to each other.

A year ago everyone obsessed over prompts. Now the real work has moved outward — building good loops, tools, and memory.

Prompt = the words
Context = what the model sees
Harness = the system it runs inside

And "pyclaudir" is an example of a harness

Amazon plans to hire 11000 interns in 2026

https://www.businessinsider.com/aws-ceo-amazon-ai-coding-jobs-interns-hiring-2026-5