Always Future Agents

I’ve been interested in software agents since I came across Graham Glass’ software ‘ObjectSpace Voyager’ in 1998. The idea behind agents is software that can act on its own on behalf of it’s “owner”, much like a human agent in the sports or entertainment field.

If you’ve ever used something like Spotlight, you’ve used a local agent. Spotlight works away in the background indexing the files on your computer so that it can answer questions like “Where did I put that presentation where I mentioned ‘Bitcoin futures’?”

There are quite a few “local agents” that are useful. But what if it’s someone else’s presentation that you are looking for? What if it’s on their laptop? To be truly useful to their owners, agents must be capable of being distributed.

In 1998 Object-Oriented was all the rage, but distributed software was still a mess. If there was a lot of money riding on it, you could use CORBA. I was significantly techy back then, and even I had trouble with CORBA. Java had Remote Procedure Calls (RPC) calls by which objects could message other objects, but the whole edifice of distributed computing was fragile. There was no platform on which you could write distributed communicating agents.

Then along came Voyager. At a stroke, Voyager let you turn a Java object into an agent that could communicate with other agents wherever they were. More amazing still was that an agent running on Voyager on your machine could “hop” to another device and execute there. It took my breath away.

Sadly, it was also useless. Almost nobody else had heard of Voyager or seemed to see its potential. There was nowhere for your agents to go and nothing much for them to do if they got there. I could never see how to make real use of it. I think this reality started to bite because Voyager pivoted and became a good, boring web application server.

But for a brief moment, I saw a beautiful future of agents communicating with each other to help their users solve problems (yes, Tron made a big impression on me as a child)!

Though it faded over the years, I’ve never entirely lost that vision. It sits as an, as yet, unexplored part of Mentat. In Mentat, scripts are a first-class citizen, and I want to make it easy to create agent scripts that perform functions on my behalf. The distribution will be achieved using TupleSpaces (an overlooked concept in distributed computing).

A simple but powerful use-case could be finding answers to questions. Imagine something like this:

  • You pose a question and post it.
  • One of your agents sends the question metadata to one or more shared tuple spaces.
  • My agents are waiting for tuples matching things I am interested in.
  • One of my agents spots your question and, realising it (a) matches my interests and (b) meets my priority requirements, ‘takes’ it.
  • It presents your question to me along with the related resources that I have on hand.
  • I select from among those resources to compose my answer.
  • My agent posts my answer back into the tuple space.
  • Your agent spots an answer and collects it to present it, and potentially others, to you at an appropriate moment.

Sounds a bit like posting a question to a web forum, right? Yes, but the differences have the potential to be transformative.

  • You don’t have to decide where to put your question; your agent can do that. Depending on your preferences, it might put it in many spaces and with different metadata depending on the space.
  • I don’t have to look for your question; my agent decides if it’s something I will want to respond to. Or ignore. Or maybe just file it away for some other purpose.
  • My agent can assemble resources on my behalf to make it easier to answer that question.
  • You don’t have to look for replies; your agent will assemble them. Potentially using a quality filter (oh, Matt replied, you’ll want/not want to see that) and potentially digesting answers. Your agent might just as well say, “You’re busy right now, but I suspect you’ll want to see Matt’s question; I will present it at another time”.

Since our agents are software under our control, we can determine how they work and improve them over time to better use our knowledge and attention.

For example, your agent might not be tasked with trying to answer my questions but simply to reflect, “Matt seems to be interested in topic X right now”. Indeed your agent might notify you not about my questions but with analysis of my questions. This could go in all kinds of directions.

I have many of the pieces of this infrastructure in place but can’t make progress right now. I really wish I could find someone to collaborate with on this platform. Hopefully, I still have a few years to get back to it and turn it into something real that people can use to solve problems.

Sharing our work

When I started working on Mentat back in 2018 I had in mind a kind of “total information store” that I could use for all sorts of things but often about outputs either to questions or in terms of content.

This was a reflection of the blurring of my work & life and the way that information tends to disappear into other peoples silos over time. I am interested in what I am interested in, no matter the context, and I would like to know what I know or at least what I thought at some point.

Meanwhile, Roam Research has come along and hoovered up a lot of use cases. I use Roam as a habitual note-taking environment. A light-weight TODO system, calendar, and personal CRM. I use it for drafting LinkedIn posts, blogs, and newsletters. And I’m a relatively unsophisticated Roam user (for example I’ve never written a query in anger, don’t use Roam/js plugins, and still use the default theme) and yet it has certainly come to dominate my digital life.

At the same time, I can reflect that one of Roam’s great strengths: its focus on blocks of written text (with tags and backlinks) is also its Achilles heel. You can put anything in Roam but structure appears only sporadically and with effort. How can you act upon what Roam knows?

In the context of, say, writing an article it works well. But what if I wanted to see if I could answer a specific question relating things that I know. That could be a lot more tricky.

Mentat comes from the opposite perspective. It deals with structured ‘things’ (indeed the root of the Mentat mental taxonomy is something called Thing). We can have a Thing called Note that represents free text. This is never going to be as powerful as Roam but, at the same time, we know what is a note and what is a instead a Person. Roam can approach this through the use of tagging. I routinely add Tags: #Person to people I add to Roam as part of my CRM but it’s not the same thing.

As yet, Roam provided few tools to act on this and of course it relies upon my consistently tagging people — which often I forget — and applying the same schema over time (mine has changed 3 times as the advice has changed). Again there are solutions to these problems but they are always a compromise of being based upon free text. Mentat has it’s own compromises but, similarly, strengths.

Three things I see as being very important to using a future version of Mentat for work are:

Being able to structure questions with appropriate metadata that allows them to be shared and acted upon by others.

A “shared space” in which questions can be placed and taken.

The ability to create agents that can act on things in the shared space. Taking them, acting upon them, changing things locally and potentially placing things back into the shared space.

Roam is going to have to tackle the problem of people sharing their graphs. That is going to be a hard problem. Mentat will allow people to create shared spaces and exchange information without needing to create a total mapping.

It will be interesting to see if (a) I can build this, (b) if it might work as well as or better (for some problems) as what Roam will come up with.

The curse of craftsmanship

I’m not a computer scientist; I’m not a software engineer; I am a programmer. I see the difference as being between a theoretical view, an engineering view, and a “craftsman’s” view of creating software to run machines. That is to say; I don’t build software, I craft it. Everything is unique, hand-made. Tools matter to me in so far as they let me express my craft. I want the tools that best help me do the work.

Here is a list of languages that I have used to solve real problems. It does not include those I learned purely for the fun of it.

  • 80×86 assembly language
  • C
  • C++ (although the C++ of the ’90s, exceptions were new back then)
  • Java
  • Perl
  • Javascript
  • Usertalk
  • Ruby
  • Objective-C
  • Clojure
  • Elixir

The choices in bold are those languages I’d go back to reasonably happily. Here are languages I actively avoid:

  • Python
  • Javascript
  • PHP
  • Go
  • Swift
  • Scala

My reasons here, as much as anything else, are reasons of taste. Python I dislike because I don’t think significant white space is a good idea (as opposed to programming in an outliner which works well and I’ve no idea why the Python community never cottoned on to this) and I never liked its feel.

Go, Swift, and Scala, are on the list because they make choices I find distasteful. For example, Swift feels like the authors have to include every possible good idea from the programming field. I’m reminded of this by Antoine de Saint-Exupery “Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away.”

More language authors should take a stand like Rich Hickey and José Valim by having a language philosophy dictating that they reject good ideas because they don’t fit that philosophy. Their languages feel coherent, and designed toward a purpose. They choose not to include the kitchen sink.

The reason I am thinking about all of this is Javascript, a language I have used to solve real problems yet, given a choice, actively avoid.

I’m sure there are good things about Javascript if you look hard enough. It’s near-ubiquity makes it a winner, and I wouldn’t criticise another’s choice to use it (esp. if they are a competitor). But, improve as it may, Javascript has cruft encoded into its DNA, and that cruft leaks out in every direction.

For a new project, I am learning Elixir and its web framework Phoenix. And do you know what has been the most challenging thing? Webpack. Worse yet, Webpack is a tool written in Javascript (using nodejs) for managing project Javascript!

Whenever nodejs hoves into view, I tend to find myself waist-deep in a mess of dependencies and version incompatibilities. Perhaps I have been lucky, but I don’t experience this level of pain elsewhere. So it makes me think it’s a function of the choice of language and the kinds of people who are aligned with those choices.

I guess it is the craftsman’s curse: You come to care about the way the tools allow you to express the work and, therefore, feel pained when forced to confront tools that feel wrong.