3,750 karma · joined September 13, 2007
If you're looking for the human touch, what about trying to offer a 15-minute zoom call instead if they want to discuss something with you? That way you'll know they're human (at least for now) and you get to connect about a shared interest.
I agree that it's a problem. I think shouting into the void isn't going to change anything. It's just going to get worse. It's similar to how I never answer my phone anymore due to spam unless the call is from a saved contact...and even then it's not 100% reliable.
(B) Advertisements in newspaper or professional journals.
(1) Placing an advertisement on two different Sundays in the newspaper of general circulation in the area of intended employment most appropriate to the occupation and the workers likely to apply for the job opportunity and most likely to bring responses from able, willing, qualified, and available U.S. workers.
(2) If the job opportunity is located in a rural area of intended employment that does not have a newspaper with a Sunday edition, the employer may use the edition with the widest circulation in the area of intended employment.
(3) The advertisements must satisfy the requirements of paragraph (f) of this section. Documentation of this step can be satisfied by furnishing copies of the newspaper pages in which the advertisements appeared or proof of publication furnished by the newspaper.
(4) If the job involved in the application requires experience and an advanced degree, and a professional journal normally would be used to advertise the job opportunity, the employer may, in lieu of one of the Sunday advertisements, place an advertisement in the professional journal most likely to bring responses from able, willing, qualified, and available U.S. workers. Documentation of this step can be satisfied by providing a copy of the page in which the advertisement appeared.
> I like the idea of using a “local inbox for agent sessions” because switching between different contexts can be more frustrating than the actual time it takes the agent to run.
> But I’m curious about how you see the inbox being useful beyond simply storing and replaying conversations.
> A few things would help me understand the practical value:
> What does a session actually contain? Is it just the conversation, or does it also include things like prompts, code changes, test results, Git commits, and other context? Can you reliably replay a session? Ideally, someone should be able to see what task was given, what the agent changed, what tests were run, and whether the task actually passed. What happens when work gets interrupted? If I pause one task and start another, can I come back later without losing the context or putting the project into a confusing state? How do you prove the work was completed? For example, can the system show the tests that were run, the changes made, and whether they passed? How is everything kept safe? If everything runs locally, that's great, but I'd still want to know how filesystem, network access, and logs are controlled.
> Even a small example showing something like how long it takes to go from starting a task to getting a working result, compared with the current workflow, would make the benefits much easier to understand.
---
Thanks for the questions! I like your suggestion about providing a small example. I'll put that together. Email me at matt@alcubi.ai and I will send it to you when it's ready. In the meantime, this page from the documentation may help to provide some of the details you're looking for https://alcubi.ai/delegator/docs/dev/concepts/ticket-lifecyc...
A session includes the scoped task request (ticket) and the code changes that are made on a separate git branch in a new worktree. You can enter the agent session at anytime with the "dg chat" command. That will allow you to look into the details of how the agent accomplished its work. So even though the runs execute in the background, the agent session is always available if you want to jump into it. I'll use this command when something was unclear to ask for follow-up changes. The agent will still have all the context in its session from the work it previously did so there's no extra context loading needed in these case.
If works gets interrupted, the ticket status moves to the Failed inbox. You'll see that clearly from the output of the "dg" command which shows the various inboxes. When this happens I'll use "dg chat" to attach to the agent session to understand what went wrong. Usually, this is something like the Claude/Codex API being down. When the API is back up, I run "dg restart" to restart the processing from where it left off. Since all work is done in separate branches and worktrees there's no ability for your project's main branch to be left in a confusing state.
The agents will run your test suite that's setup. delegator is just the orchestration layer. You can customize any steps you'd like the agents to follow via your AGENTS.md file. That will be taken into account when the agents run. So for example, you could add in your AGENTS.md file that for any delegator ticket they work on to include details in the commit of tests that were run. It's very flexible like that so you can customize it for your setup.
For safety, I've gotten to the point where I trust that agents are only accessing data that they need so there are currently no guardrails in place. This is something you could add to your AGENTS.md file to get information about but that's not deterministic. I could add in features to support this. Please email me at my email above to let me know a bit more about what you'd like to see and I can work on that.
Hope this helps and thanks for taking a look!
> Do you keep a fresh context per agent?
Yes I've run into the same problem of trying to stuff too much into one session. So Delegator uses a fresh session for each task. For larger workflows, I'll do the same where I'll have a design or feature doc and then the tickets may refer to that if necessary for additional context.
> The part that eats my time isn't the delegation, it's the review. I trust an agent now on mechanical work with a crisp "done" definition, and a lot less where the acceptance criteria are fuzzy, so those tickets tend to take me longer than doing them myself. > Between the queue and your review, where do you actually get bottlenecked? That's my constraint right now.
Yes, I'm bottlenecked on the review. I think everyone has this problem now. I'm trying to solve that by bringing the task into a ticket system so you can easily look up what it was trying to do. From some research I did, context switching isn't so bad if you are presented with the context clearly and succinctly when coming back to the task. I'll admit I don't think I have a perfect solution for this yet but I'm trying to work towards that. I do think this is the biggest issue with AI-assisted development - reviewing the code and output. What have you tried that's been working?
Have there been cases where kids in the 95% did end up being self-driven about a particular topic that they had an interest in? Or did they always need to be given motivation?
> It is therefore the policy of my Administration that, to the maximum extent permitted by law, the executive branch shall use the terms “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” and will not acknowledge the usage of “Artificial Intelligence” and “AI” in any applicable setting.
AI will help advance the speed of progress but they won't be autonomous.
Models will continue to get more optimized and you'll be able to run what you need locally on your computer.
It's an interesting question. The thing I keep coming back to though is that every time I've tried to go more towards vibe-coding, I invariably look at the code and find things have been added that would just not be acceptable. I've also tried asking the models to see could be refactored however they still miss things that should be obvious.
I think the gap is that they're still lacking a sense of importance. As engineers working on a product, you have a sense that this feature is more important than that feature. An LLM treats your codebase at the same level of importance. So they'll spend the same amount of effort and code changes on testing and hardening something that just really isn't that important.
Also, once a bad pattern gets into the codebase, they just continue to build and extend that out rather than re-thinking about it like an engineer would.
I've tried this out. Even with relatively small applications and with spending hours reviewing the spec documents, there was always something I missed or something that wasn't quite right when seeing it live.