do not look into laser with remaining eye

The fourth in a series; see 1) considerations, 2) refinements, and 3) design.

There's been a number of notable cyberattacks perpetuated by "independent" AI agents.

Building on the previous posts (considerations, then refinements), this post outlines a design for a multi-agent team.

Update: this is the second in a series; see the subsequent design.

I've written an agent that uses API calls to an LLM and local tools.

I enjoy conversation that has a high rate of insights.

Software projects typically have a folder structure like /project-name-here/ ├── src/ # Source code │ └── lib/ # Libraries or modules ├── test/ # Unit tests, regression tests, ├── docs/ # Documentation ├── configs/ # Configuration…

Teaching folks how to use AI includes conveying the scope of what aspects of problem solving can benefit from use of AI.

Ignoring latency, how is one agent taking on distinct roles with various tools different from a multi-agent collaboration?

In a previous post I proposed a workflow (akin to https://en.wikipedia.org/wiki/V-model) of sequential tasks that could provide the agent and the human clear stages of progress for software development.

Building a team requires filtering people who would harm the organization, whether due to incompetence or laziness or something else.

Not all sources of second-hand information should be treated equally, and not all audiences are equivalently dangerous with second-hand information.

There are a lot of popular websites that rely on personalized front-pages.

I have an LLM agent written in pure python that has access to bash commands.

One-shot prompts to get a result are one way of engaging LLMs.

The following is a math problem that I encountered.

Faced with the need to justify allocation of resources, how does a person who doesn't understand the decisions that were made respond?

Every organization that relies on digital data faces a Kodak-like transformational choice: how to invest in leveraging LLMs versus continuation of legacy processes.

I asked Gemini (Google's LLM) for a description of a topic and process I had heard of but wasn't familiar with.

Identifying things that should be better is relatively easy; understanding the forces that act against change is an important step towards enacting change.

One of the nuggets in Doctorow's Code is a liability (not an asset) is a reminder that sequential LLM's (e.g., in an agentic workflow) cause system success to decrease.

In my experience with using LLMs for both simple and complicated topics, I have observed that having general awareness of a topic's existence and (when possible) deep expertise in that topic are both useful.

Currently we expect LLMs to be correct.

This post was last updated 2025-10-09.

Large complex systems (which often don't work well and are inefficient) evolved from smaller systems that worked.

Bureaucracy involves a chain of two principal-agent problems: - Society (principal) has shared resources that are entrusted to policy makers (agents). and - The policy-maker (principal) puts a bureaucrat (agent) in charge of enforcing the policy for the s…

In a livelock condition, "the states of the processes involved in the livelock constantly change with regard to one another, none progressing." source: https://en.wikipedia.org/wiki/Deadlock_(computer_science)#Livelock Similarly, bureaucratic livelock is…

A provider of LLMs spends money on hardware, electricity, data centers, researchers, managers, HR, and lobbyists.

I had PNGs of a book that I used Gemini 2.5 LLM to OCR and transcribe plain text.