Jonathan Abend

You are right. You could curate the data for continuous learning as well.

I have always thought of continuous learning as automated retraining without any dataset curation. But who says that you can not do that.

I don't see the point / the use cases of continuous learning. What does it solve what context can not solve?

In traditional data science or narrow AI, you had to curate your training data very carefully. That's why the thought of letting models learn from 'random' data seems strange to me.

Next steps:

- publish it
- runtime envs via docker compose
- cloud storage as workspace
- support for open source LLMs
- support for other cc-like frameworks
- company-like hierarchical governance features for keeping track of what the company of agents is doing 24/7
(e.g., reporting, alignments..)

Each agent rates its work environment after each run.

It provides suggestions for
improving the setup and rates the quality/efficiency of received tasks, work results of
other agents and tool usage.

This helps users to find the spots where they can improve the agent configs.

Users can view the logs, input and output of the CC instances within the menu popups
for each agent node. Over there, you can also configure the agent (CLAUDE.md file, timeout etc.)

A central flow control invokes independent CC instances via Claude Agent SDK.

Currently, the agents can either communicate via prompt templates or Chainlink issues

Thanks for the helpful tool @dollspace.gay

I am working on a tool for orchestrating multiple Claude Code instances in more
company-like structures

(e.g., setting up a continuous loop between developer agent
and a product owner agent which tests and validates).

While Claude Code and I are having the time of our lives,
my anti-virus scanner is annoyed and keeps deleting uv from time to time.
I suspect it deletes git and python as well sometimes.

Orchestrate teams of Claude Code sessions - Claude Code Docs Coordinate multiple Claude Code instances working together as a team, with shared tasks, inter-agent messaging, and centralized management.

Very interested to see how this goes

code.claude.com/docs/en/agen...

MCP Apps - Model Context Protocol Build interactive UI applications that render inside MCP hosts like Claude Desktop

In case you were also wondering what MCP apps are:

'Sometimes users need to interact with data, not just read about it. MCP Apps let servers return interactive HTML interfaces (data visualizations, forms, dashboards) that render directly in the chat.'

share.google/g8ejY3xhnO8d...

#MCP #AI #data

The design of today's corporate processes and hierarchy levels is based on cognitive limitations of humans.

Is Your Workplace Set Up for AI Agents? AI’s true productivity gains require redesigning organizations, not merely adding AI to human-centered systems—much like factories once had to redesign around electricity. Current productivity estimat...

Great piece on how companies need to transform to fully leverage the potential of AI agents:

hbr.org/2026/01/is-y...

1) Reduce data silos

2) Markdown besides human-centered file formats like ppt

3) One API/MCP-based interface next to each human-centered interface

4) agent-to-agent processes

Nice finding that the performance of two agents is worse than the performance of one agent when the agents have the same role (in this study developer).

However, I see the value of multiple agents anyway rather in setups where each agent has its own scope, e.g., 1 dev agent, 1 product owner agent.

When building AI agents,

let them use bash, when you would like to get trial-and-error-like exploratory problem solving

and tools, when you want to enforce a certain way of acting (e.g., in sensitive environments) and observe these actions.

Need to try that

Claude in Excel is really good.

Its weird that using Microsoft's own Excel agent using Claude 4.5 often yields weaker answers, but it seems to be because the Excel agent relies on Excel alone (VLOOKUPs, etc) while Claude in Excel does its own analysis and uses Excel as a UX for input/output.

In our new world of LLM-assisted programming, Test-Driven-Development no longer is a "nice to have" workflow.

It now is essential.

Demand tests for everything: basics, edge-cases, integration ...

Use proper adversarial reviews and create more tests before fixing the code accordingly.

Just do it!

Claude Code costs up to $200 a month. Goose does the same thing for free. Goose, Block’s open-source AI coding agent, is emerging as a free alternative to Anthropic’s Claude Code, as developers weigh offline control, rate limits, and the rising cost of AI coding tools.

aaaay I missed goose in the news

I am using hooks for running automated unit tests, linting, playwright etc.

I haven't tried out Ralph because I am a bit scared that it burns a lot of tokens. But a well-formulated definition of done is probably helpful here as well.

Have you tried it? Does it work for you?

Claude Code - Ollama

Ooh. Ollama just released Anthropic API impersonation, meaning you can set the endpoint to a local port and be off to the races! docs.ollama.com/integrations...

I definitely need more hacks to keep Claude Code in the loop for longer.

For example:

Instead of telling it "Fix code A which does B",

I am now telling it "Please do B by using code A. If you run into issues: fix them until you can return B to me."

Proof of work / result artifacts are key.

I started it to give CC the chance to focus more on single tasks and dig deeper into details.

If I am overwhelmed, CC might get overwhelmed, too.

Letting Claude Code create Github issues and epics based on the plan it has created feels amazing (slash command 1).

Letting CC solve these issues one by one feels even better (slash command 2).

Claude Code: Part 1 - The Foundation Building a sophisticated Claude Code configuration system with smart wrappers, mixin architecture, and orchestrator-only discipline.

Out-of-the-box Claude Code is great for simple stuff. For complex projects it'll drown you in AI-generated mediocrity unless you configure it properly.

Wrote up my setup. Part 1 of 3:
rungie.com/blog/claude-code-foundation/

In 2026, the differentiator won’t be the LLM—it’ll be:
— who owns the context
— who integrates memory
— who designs with guardrails
Next-gen AI is less about size, more about fit.

#AI #GenAI #LLM

YouTube video by Google Cloud Tech Identity and Access Management for Agents

Nice explanation of how to establish a secure identity access management for AI agents: youtu.be/DQX81oJfsTE?...

Define user/admin roles to access/develop (specific) agents.

Let the tool code handle data access. The agent only passes a session ID. The tool fetches secrets from a secret manager.

Amazon Q - Visual Studio Marketplace Extension for Visual Studio Code - The most capable generative AI–powered assistant for software development.

One of the next things on my to do list is to try out the Amazon Q extension for VS Code. Saw it last week on the screen of my colleague.

Amazon Q - Visual Studio Marketplace share.google/yYptrSt7WhIK...

Quite impressive and thoughtful what Google's Vertex AI offers for agent development and #AgentOps

Pitschnass

Today, I tried to enhance an existing application with the Cline extension for VS Code. (It's like vibe coding with Cursor - just in VS Code). But I am not convinced of the results yet.