amthewiz

Agent Applications are AI systems built around persistent, tool-using agents.

I'm the author. Working draft 0.9. The claim is that "agent application" should be a recognizable shape the way "web application" is - the paper covers shape of an agent application project, how it gets deployed, how instances map to privacy boundaries, what belongs in the durable workspace, and how fleets of diverged instances get upgraded.

Two things I'd genuinely like from this crowd: Have you tried building an agent application from the current pieces - harness, MCP, skills - and thought the process was arbitrary? And how do you run harness-based agents in production today, where each one gets a filesystem, command line, and MCP tools - Claude Managed Agents, Deepagents, something else?

(Disclosure: I build karta.sh, one harness-neutral implementation of the platform layer the paper describes. The paper cites it alongside Bedrock AgentCore, Vercel Eve, and Pickaxe.)

It has puzzled me why someone hasn't already done this yet, given LLMs are so good at language now.

Probably the short answer is that is it hard to get this to actually work. There were many open questions that one has to tackle simultaneously -

- What is the right balance between relying on LLMs to do the right thing vs the runtime around LLMs? For example, I went back and forth a few times getting LLMs to manage stack as you call one playbook from another one. Finally decided that it is most reliable to let the runtime take care of that.

- Context engineering - What to put in the prompt, in what order, how to represent state, how to handle artifacts, how to make sure that we use LLM cache optimally as context grows, how to "unwind" context as playbook calls return, how to compact specific types of information, how to make sure important context isn't lost, etc

- LLMs today have vastly different capabilities than 2 years ago. I have had to rewrite the whole stack from scratch 4 times to adjust. Wasn't fun, but had to be done.

- Language(s): How to represent the pseudocode so that it is both fluid natural language and a capable programming language? How to transform it so that it can be executed reliably through LLMs? How to NOT lose the flexibility and fluidity in the process (e.g. easy to convert to a graph like LangGraph, but then you are stuck with control flow), how to create a semantic compiler for that transition, what primitives to use for the compiled language that I call Playbooks Assembly Language [1].

- Agents and multi-agent system considerations - How to represent agents, how they should communicate. Agents are classes and they expose public playbooks that other agents can call). Agents can send natural language messages to each other and engage in conversations. Agents can call multi-party meetings. How can the behavior across all these interaction patterns be defined so it remains intuitive. For example, lifetime of a meeting is tied to "meeting: true" playbooks so the agent simply returns from the playbook to exit a meeting and meeting lifecycle is tied to the host returning from its meeting playbook.

- Which LLMs to support? Go for "Bring your own LLM" or restrict the set? Which LLM? LLM selection impacts how all the internal prompts are implemented so prompt building had to occur simultaneously with LLM selection.

It felt like playing an N-dimensional game of chess!

[1] https://playbooks-ai.github.io/playbooks-docs/reference/play...

Andrej Karpathy posted in early 2023 (https://x.com/karpathy/status/1617979122625712128) -

> "The hottest new programming language is English"

I've built a Natural Language Programming stack for building AI Agents. I think it is the first true Software 3.0 stack.

The core idea: Use LLMs as CPUs! You can finally step debug through your prompts and get reliable, verifiable execution. The stack includes a new language, compiler, developer tooling like VSCode extension.

Programs are written as markdown. H1 tags are agents, H2 tags are natural language playbooks (i.e. functions), python playbooks. All playbooks in an agents run on the same call stack. NL and python playbooks can call each other.

Quick intro video: https://www.youtube.com/watch?v=ZX2L453km6s

Github: https://github.com/playbooks-ai/playbooks (MIT license)

Documentation: https://playbooks-ai.github.io/playbooks-docs/getting-starte...

Project website: runplaybooks.ai

Example Playbooks program -

    # Country facts agent
    This agent prints interesting facts about nearby countries

    ## Main
    ### Triggers
    - At the beginning
    ### Steps
    - Ask user what $country they are from
    - If user did not provide a country, engage in a conversation and gently nudge them to provide a country
    - List 5 $countries near $country
    - Tell the user the nearby $countries
    - Inform the user that you will now tell them some interesting facts about each of the countries
    - process_countries($countries)
    - End program

    ```python
    from typing import List

    @playbook
    async def process_countries(countries: List[str]):
        for country in countries:
            # Calls the natural language playbook 'GetCountryFact' for each country
            fact = await GetCountryFact(country)
            await Say("user", f"{country}: {fact}")
    ```

    ## GetCountryFact($country)
    ### Steps
    - Return an unusual historical fact about $country

There are a bunch of very interesting capabilities. A quick sample -

- "Queue calls to extract table of contents for each candidate file" - Effortless calling MCP tools, multi-threading, artifact management, context management

- "Ask Accountant what the tax rate would be" is how you communicate with other agents

- you can mix procedural natural language playbooks, ReAct playbooks, Raw prompt playbooks, Python playbooks and external playbooks like MCP tools seamlessly on the same call stack

- "Have a meeting with Chef, Marketing expert and the user to design a new menu" is how you can spawn multi-agent workflows, where each agent follows their own playbook for the meeting

- Coming soon: Observer agents (agents observing other agents - automated memory storage, verify/certify execution, steer observed agents), dynamic playbook generation for procedural memory, etc.

I hope this changes how we build AI agents going forward for the better. Looking forward to discussion! I'll be in the comments.

In the blog, they say

> we’re launching the GPT Store, featuring creations by verified builders

Any idea who these “verified” builders are, how they are selected and how to submit a GPT that gets selected in the rollout?

Children can also feel pleasure and love.

The field of software engineering changes rapidly and is complex to master. To remain good at the job, one must continue to explore and learn. The only way that is going to happen is if you are genuinely interested in the field. The best indicator of that is that you have pet projects outside of work.

> So Japanese people are doing everything right -- eschewing violence, avoiding drugs, working hard and not having kids out of wedlock. They are following the conservative prescription, as well as or better than any other developed country in the world. And yet still, many of them are poor. This suggests that there is something very wrong with the conservative theory of poverty.

No, the author is confused. There is a big difference between -

- you cannot be poor if you make good choices, and

- you can get out of povery by making good choices

The conservative position is the later. There will be cases where it doesn't work, but it is a sound recommendation in general.

Now, poverty does make it harder to make good choices. So welfare $s should be spent to make that easier, e.g. free food at school helps kids do better at school.

Basics should be concepts that get you to 80% and tell you where to look for the rest 20%. This book tries to get you directly to 95% and is best treated as a reference book.

Who does that?

People go their entire lives without knowing what it is like to go without food for a day. They take it for granted.

Also, I believe parents generally don't teach kids to respect food. It probably starts by letting a toddler feed him/herself too early and in the process normalizing food wastage.

There is also a first world etiquette of leaving some food in the plate. I have seen people habitually do it.

The DNN based techniques are new but the concept of models that can fill in the blanks is old. It used to be called content addressable memory or autoassociative memory.

I'd love to take the robo-callers on a long winding road to nowhere. Is there a conversational AI robo-answering app based on something like Google Duplex? If not, should I build one?

Agreed. The title is doing disservice to the content.

Tuche!

An advanced civilization, when looking to signal other potential civilizations, would hopefully -

- use the lowest common denominator long range communication medium (like EM waves), not the most advanced technology available to them, and

- would broadcast it widely in spatial, temporal and spectral dimensions and not worry about efficiency

Really? I don't like to fetch water from a lake every day to my house. Instead of using technology to bring water to my house, what lifestyle changes or hard choices should I rather make?

Are they sincerely suggesting that the neural correlate of consciousness is a new type of neuron? I hope not.

To go even further on that point - the chart itself seems overly optimistic. +-40% of estimate after product design? +-10% of estimate after design spec?! Countless number of times I have seen detailed design specs completely thrown out of whack by new engineering, process or business revelations/insights. Human thought is iterative - most people can imagine at most 80% of the future/product/design possibilities.

What would be the best way to help sci-hub anonymously if one wishes to do so?

Current neuroscience suggests that brain operates in broadly two regimes - learning during wakefulness and model selection during sleep. Both these activities require "forgetting" of specific kinds. Specifically model selection involves trimming synapses. The goal is to learn a model of the world at the right level of detail, what AI folks call avoiding over fitting.

As I understand, the basic idea is to minimize surprise over the organism’s life using models of the world to predict the causes behind what the organism observes and actions to keep the organism within states suitable for survival.

Ruby's optional parenthesis for function calls. Makes beautiful DSLs like RSpec -

  describe UsersController do
    before do
      allow(controller).to receive(:current_user).and_return(nil)
    end

    describe "GET #new" do
      subject { get :new }

      it "returns success" do
        expect(subject).to be_success
      end
    end
  end

I see what you did there.

Current state of the art precision has yielded insights at the sensory-motor (e.g. precise understanding of what sensory inputs cause what neural activity) and at subsystem level (e.g. what mental activity "uses" what brain subsystems). I am sure more can be done with existing technology, but a grand unified theory would probably need more precision. It is somewhat akin to attempts at finding the grand unified theory of Physics that works both at quantum and cosmological scales.

> would you say that your dreams are observations of physical events?

Current understanding is that dreams are a side effect of variational model selection. See Hobson, J. A., & Friston, K. J. (2012). Waking and dreaming consciousness: Neurobiological and functional considerations. Progress in Neurobiology, 98(1), 82–98. https://doi.org/10.1016/j.pneurobio.2012.05.003

> brain scans of subjects under these conditions find anything interesting

It is certainly possible, but currently not practical. Imaging technologies have limited spatiotemporal resolutions and coverage areas. Direct electrode reading are too intrusive, and so on. There is limited understanding of which brain areas might contribute to certain complex behavior so we might have to measure entire brain's activity. We don't know what resolution is useful - each spike or firing frequency or only dynamics (changes), etc. If we assume the availability of whole brain scan at spike resolution over extended time periods, theoretically a regression type algorithm could figure out relationship between brain activity and behavior. Firmly out of currently technology's reach though.

> definition/enumeration of self-awareness

Philosophers have been working on that for a long time. Some terms are well settled now, even though a term might cover a range of concepts. See that for "consciousness" here - https://plato.stanford.edu/archives/spr2018/entries/consciou.... This addresses mostly the western thought process and leaves out the eastern concepts. Most eastern concepts are too vague to pin down, though some are concrete enough to provide some insights (e.g. http://heavenmeetsearthyoga.com/news-features/the-meaning-of... matches reasonably well with modern understanding).

The conference sounds like a nightmare for serious scientists and philosophers. Consciousness is a broad term with lots of nuances - (https://plato.stanford.edu/archives/spr2018/entries/consciou... presents a good summary). Following unconventional and novel approaches to understanding consciousness are fine, but the Deepak Chopra style dancing on the linguistic vagueness is infuriating. I am glad I did not attend this conference and instead opted to present some ideas at ASSC conference later this month.

This.

Would people be willing to shell say $5/mo towards their online news content consumption?

A browser plugin could then track time spend (keep all data local), at the end of the month, present a simple dialog that says, your top 5 destinations were these, uncheck any that you don't want to pay, and rest will get your $5 divided by time spent.

This way users would get to browse ad-free, thanks to the ad blocking solutions, while publishers would also get some revenue. Ideally the payment service could generate some revenue as well and pass some to further developement/maintenance of the ad blockers.

I'd recommend OnePlus 5T or the upcoming 6. Switched from iPhone a few months ago and happy.

My 2012 MBP is still going strong and I see no need to replace it. That never happend to me with Windows laptops that I used for a decade before that.

I love the trackpad of MBPs. The aesthetics and physical build are good. The OS hardly ever gets in my way. The UI has been stable and usable and improving over the last 6 years of use. Apple's hardware service has been great (had video card issue a couple of years ago - $200 flat rate fix got me new motherboard, new screen with top cover, new video card). It has all the ports I need. Magsafe connector is genious. The display is crystal clear and high res.

I have had a genuinely good experience with my 2012 MBP. Barring bad hardware luck and missteps by Apple like the touch bar, people complaining about MBPs are being nitpicky.