Ignoring latency, how is one agent taking on distinct roles with various tools different from a multi-agent collaboration? How does having two or more agents add value over having a single agent? Answer: overlapping context windows that are concurrent. Compare the two situations. First, a single LLM sequentially taking on two roles must switch contexts. Figure 1: One LLM taking on sequential…
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. There are some clear benefits, like ending up with a list of use cases and requirements and design decisions. However, this prescribed sequence didn't work well in practice when I tried to use the…
Building a team requires filtering people who would harm the organization, whether due to incompetence or laziness or something else. As an interviewer, you might have 30-60 minutes to determine how your organization could benefit from or be harmed by this candidate. The post provides interview questions to inform that decisions. Before getting to the interview questions, it is important for you…
Not all sources of second-hand information should be treated equally, and not all audiences are equivalently dangerous with second-hand information. If the provider and/or receiver has out-of-band context, the second-hand information is more likely to be better conveyed. If the person giving or receiving the second-hand information lacks the context, then there will be confusion and…
There are a lot of popular websites that rely on personalized front-pages. Examples includes Reddit, YouTube, Facebook, LinkedIn, Nextdoor, Instagram, Tiktok, Twitter. The "for you" content is more harmful than the secondary effect of an echo chamber. The echo chamber, while a valid concern, is an accidental outcome. Whereas the tailored content is engineered to keep you on the platform to consume…
I have an LLM agent written in pure python that has access to bash commands. The structure of the script that creates the agent is # system prompt identifies for the LLM which bash commands are available while True: # Get user message. (This is where the human interacts for input.) # append user input to context while True: # query LLM to process context if bash_command_in_LLM_response: # seek…
One-shot prompts to get a result are one way of engaging LLMs. Iterating a session with multiple prompts is a way to refine the nuance or explore related areas. This post focuses on orchestration of prompts which break a complicated task into a sequence of smaller, focused steps where the output of one prompt becomes the input for the next. Prompt orchestration encodes best practices for…
The following is a math problem that I encountered. Situation : There are four buckets being concurrently filled with water. When a bucket is full it tips over and empties. Parameters : Each bucket has a different fill rate: 10, 12, 14, 16 seconds. Emptying each bucket takes 1 second. Question : How frequently do all four buckets empty at the same time? Answer: Imagine all the buckets emptying at…
Faced with the need to justify allocation of resources, how does a person who doesn't understand the decisions that were made respond? One approach would be to build relationships with stakeholders, gain competence, and apply judgement. However, those all take time and don't scale well. Another approach is ignorance-based mimicry which leads people to enact bureaucratic tendencies: forms,…
Every organization that relies on digital data faces a Kodak -like transformational choice: how to invest in leveraging LLMs versus continuation of legacy processes. Here "organization" includes commercial businesses, government, and non-profits. The non-physical aspect of digital data empowers LLMs, and many organizations rely on digital data. The reliance on data is separable from tangible…
I asked Gemini (Google's LLM) for a description of a topic and process I had heard of but wasn't familiar with. Over course of about 6 one-shot questions I was able to learn what I was interested. For comparison, and to see if I could find any hallucinations in what I had gotten from the LLM, I reviewed the Wikipedia entry on the topic. The wiki content was not tailored to what I was seeking and…
Identifying things that should be better is relatively easy; understanding the forces that act against change is an important step towards enacting change. What should be accounted for when attempting to enact change? incentives of individuals involved. motives of individuals what narrative is told of the future of the past your reputation who controls resources needed for change who has power who…
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. If two sequential LLMs are each right 95% of the time, the cumulative correctness is 0.95*0.95=0.9025. That doesn't bode well for agentic systems, but then again what is human society if not a set of interdependent agents? No one…
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. Having a general awareness for a topic helps with knowing what questions should be asked and how to frame the prompts to get the most insight. Without general knowledge I would be less able to…
Currently we expect LLMs to be correct. When there are false statements we say the LLM is hallucenating. Hallucination is the wrong framing as it implies 1) LLMs are thinking and 2) the thinking defaults to being correct. The more accurate framing is that we have trained machines to dream, and sometimes the dreaming is lucid .
This post was last updated 2025-10-09. This post is scoped to text generation. The advice below should be augmented by you (the reader) learning about what LLMs are and how they work. This knowledge decreases the number of assumptions about what is and is not possible. The advice below is unordered. Prompt writing Prompts without context will have low-quality responses. Provide as much detail as…
Large complex systems (which often don't work well and are inefficient) evolved from smaller systems that worked. Q: Why not just stick with "small" processes that work well? A: Throughput, complexity, cost. Throughput: try processing all taxes from every citizen with a small process. Complexity: try building a modern airliner with a small process. Cost: The cost of bespoke interactions is higher…
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 shared resources. For each of these the principal is often dissatisfied with the agent's actions. Alignment and accountability are enabled by…
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 when participants in a process are busy but nothing productive results from the activity. This can arise when (duration for a process to…
A provider of LLMs spends money on hardware, electricity, data centers, researchers, managers, HR, and lobbyists. What are the options to generate revenue? Sell subscriptions for access to better LLMs Sell subscriptions for early access to new LLMs Enact rate limiting queries to drive users to subscription model Generate ads tailored to the user's search history. Sell the service of automating the…
I had PNGs of a book that I used Gemini 2.5 LLM to OCR and transcribe plain text. I then used Gemini 2.5 LLM to convert the plain text to Cypher for Neo4j. That seems preferable over the GEDCOM syntax. Next problem was visualizing the ~350 nodes and relationships in a hierarchical structure. Neo4j Bloom - https://neo4j.com/product/bloom/ Neo4j save view as SVG networkx's ancestors -…
Step 1: pictures of text My uncle wrote a genealogy book. I took pictures of the pages. Step 2: images to text I used this prompt to convert images to text prompt transcribe these images. keep the indentation consistent A rate limit on the output: even though I could upload the 38 pictures, the output was constrained by Google (to something like 8000 tokens?). So I ended up batching 6-7 images per…
https://en.wikipedia.org/wiki/G._K._Chesterton#Chesterton's_fence can be summarized as, "If a fence exists, there is likely a reason for it." Therefore, when creating a new policy, document that reason. The reason might be an anticipated problem, a near-miss, or an incident. Policy documents might state a goal, but they rarely provide the triggering event. So why not include the motive in…
Observation 1 : Communication among people within an organization are dependent on the distance between people. (That's called the Allen curve and was developed in the 1970s when in-office work was the norm.) Observation 2 : When an organization of people creates a product, the design of the product is consistent with the communication of the participants. (That's called Conway's Law .) Therefore,…
The phrase "Scratch my back and I'll scratch yours, or else I'll scratch your eyes out" vividly captures the interaction with coworkers. Understanding the dynamics can help you improve your effectiveness. This post identifies what coworkers can do for synergy (positive) and punishment (negative). The utility of this is to consider "what can I do better to contribute to a positive working…