Very cool setup. I went down a similar path and while it is definitely cool to see the extent of how agentic AI can be, I agree the ROI isn't _quite_ there yet.
I don't live in a world where I need to be constantly reading and replying to emails, so 95% of my inbox is just subscription spam. With development I need to be at the helm to design the planning requirements and actively make decisions before the agent goes off and executes the plan. But I don't need a personal agent for that, I work directly out of codex/claude-code.
Skills and MCPs CAN bloat the context window. Some harnesses do more progressive loading of skills/mcp depending on how many you have. Check your harnesses docs.
Personally, I feel skills+CLI is better, since only the description of the skill enters the context window until it’s required and CLIs should support help flags which will allow you do have progressive discovery.
First, it matters how well the MCP is written; a well-written one won't be so massive
Secondly, different sessions should load different MCPs. Load only what is needed
Thirdly, if you're finding you need a dozen MCPs, you need to consolidate that into a single service that does the work and then presents a single, unified API (via MCP) to your agent
Finally, start using models with 1M context window limits. I really don't know how people can stand being stuck at a limit like 272K.
being able to talk to each of the agents via dm (but also in group chats) sounds interesting
does that mean that you have 1 chat per domain specific agent? can you also start multiple sessions/threads or is that not part of the way you interact with them currently?
it's just like Slack really. I have a DM open for each agent where I can talk direct. But they are also part of individual rooms also where I just need to @ them and they open a thread. They can talk to each other also by @'ing each other (in public rooms or ones they both belong to)
Not pointed at the author, but at the current state of affairs: this is fucking exhausting. We went and made a trillions-dollar market out of the bikeshedding maximization machine.
LLM was invented and its just clear that this is something someone needs to build.
Why?
Because it makes just sense. You don't want an agent running on a laptop you close. You want to keep context small, you want to split up work / parallize it etc.
I'm now waiting for a while until the open source agent platform emerges and im borderline motivated to build something but i'm not doing it. He did, which is not a crime.
I've been building a triage agent for my inbox and whatsapp (it's product shaped), which has ironically left me not building one of these. So even while productizing, I'm getting fomo on the full monty.
I've also been building a harness that maintains my apps which I'm hoping to open source.
Hard agree that these things don't have personal ROI, and are actually quite hard to build reliably.
But it's really fun! And having a bot fix a live error is pretty exciting.
that was one of my FAQs at the bottom. I want separation of duties and least-privilege so agents can only access what they need for their particular duties.
That said, my vision is to eventually build manager agents to manage the minion agents. Like real people. I have no idea if "people" is the right analogy for all this work but my brain can't really wrap around a different analogy yet.
I have gone back and forth on a similar sentiment a few times. I settled on this: the only responsible use of LLMs is to fill in gaps.
It's fine as a search tool or autocomplete. It can be okay to generate code if you didn't know how else to get started, or you've already limited the damage it would do by your own design.
People who overuse LLMs are usually trying to compensate for their lack of experience, structure in their work, or dysfunctional teams. Anyone trying to get hired should recognize it as a new red flag attempting to cover up the old red flags.
$48/month for the droplet (could probably be cheaper on Hertzner) and $100/month for OpenAI plan. so ~$150/month but again this could be cheaper with a different VPS and using Terra/Luna.
I also have a claude max plan for coding but like I mentioned in the article coding is still separate from the agents.
A box with 2vcpu, 4gb ram and 50gb hdd on https://shellbox.dev is $0.02/hr and you pay per minute only for what you use. Even using it 24/7 is like less than $15
$48/month for the droplet (could probably be cheaper on Hertzner) and $100/month for OpenAI plan. so ~$150/month but again this could be cheaper with a different VPS and using Terra/Luna.
I also have a claude max plan for coding but like I mentioned in the article coding is still separate from the agents.
I actually went down a different path and set it up so my friend and I can use the same agent with approval-gated turns. He's more technical than I am, but I believe I'm better at marketing. We can see both of our prompts and be on the same page. There is also a mode where we can branch off with our own agents.
Watching his prompts made me better at producing my own and using the agent effectively.
tried similar setups in the past, but messages is just not my jam, so built a html wrapper on top of claude code and open code that runs on a small-ish instance. looks like this https://i.imgur.com/lj9Fgco.png (screenshot anonymized with chatgpt) and allows multiple conversation per project, and has a few conveniences like scheduled tasks. one of the project in the list is the project itself, so I can add features whenever.
geooff_ | 3 hours ago
> Has it been worth it? For the journey, yes, for the ROI, nope.
It's also nice seeing someone experiment without succumbing to AI psychosis.
[OP] carimura | 3 hours ago
orangebread | 3 hours ago
I don't live in a world where I need to be constantly reading and replying to emails, so 95% of my inbox is just subscription spam. With development I need to be at the helm to design the planning requirements and actively make decisions before the agent goes off and executes the plan. But I don't need a personal agent for that, I work directly out of codex/claude-code.
[OP] carimura | 3 hours ago
whazor | 3 hours ago
It does not cost too much effort to maintain MCP servers. No port forwarding or VPNs required thanks to OpenAI tunnels.
And security wise its quite nice, since you have to activate MCP or give permission sometimes. So each chat is kind of isolated from each-other.
[OP] carimura | 3 hours ago
whazor | 3 hours ago
Atotalnoob | an hour ago
Personally, I feel skills+CLI is better, since only the description of the skill enters the context window until it’s required and CLIs should support help flags which will allow you do have progressive discovery.
And a CLI is also easy to use yourself.
trollbridge | an hour ago
First, it matters how well the MCP is written; a well-written one won't be so massive
Secondly, different sessions should load different MCPs. Load only what is needed
Thirdly, if you're finding you need a dozen MCPs, you need to consolidate that into a single service that does the work and then presents a single, unified API (via MCP) to your agent
Finally, start using models with 1M context window limits. I really don't know how people can stand being stuck at a limit like 272K.
tosh | 3 hours ago
being able to talk to each of the agents via dm (but also in group chats) sounds interesting
does that mean that you have 1 chat per domain specific agent? can you also start multiple sessions/threads or is that not part of the way you interact with them currently?
[OP] carimura | 3 hours ago
It's literally just like humans.. except. not.
tosh | 2 hours ago
moribvndvs | 3 hours ago
[OP] carimura | 3 hours ago
Manfrednotfunny | 3 hours ago
LLM was invented and its just clear that this is something someone needs to build.
Why?
Because it makes just sense. You don't want an agent running on a laptop you close. You want to keep context small, you want to split up work / parallize it etc.
I'm now waiting for a while until the open source agent platform emerges and im borderline motivated to build something but i'm not doing it. He did, which is not a crime.
morkalork | 28 minutes ago
JSR_FDED | 2 hours ago
lukasco | 2 hours ago
I've also been building a harness that maintains my apps which I'm hoping to open source.
Hard agree that these things don't have personal ROI, and are actually quite hard to build reliably.
But it's really fun! And having a bot fix a live error is pretty exciting.
4lx87 | 2 hours ago
[OP] carimura | 2 hours ago
That said, my vision is to eventually build manager agents to manage the minion agents. Like real people. I have no idea if "people" is the right analogy for all this work but my brain can't really wrap around a different analogy yet.
phplovesong | 2 hours ago
[OP] carimura | 2 hours ago
sublinear | 31 minutes ago
It's fine as a search tool or autocomplete. It can be okay to generate code if you didn't know how else to get started, or you've already limited the damage it would do by your own design.
People who overuse LLMs are usually trying to compensate for their lack of experience, structure in their work, or dysfunctional teams. Anyone trying to get hired should recognize it as a new red flag attempting to cover up the old red flags.
gchp | 2 hours ago
[OP] carimura | an hour ago
I also have a claude max plan for coding but like I mentioned in the article coding is still separate from the agents.
messh | an hour ago
owencmcgrath | 2 hours ago
Great write up!
davidw | an hour ago
[OP] carimura | an hour ago
$48/month for the droplet (could probably be cheaper on Hertzner) and $100/month for OpenAI plan. so ~$150/month but again this could be cheaper with a different VPS and using Terra/Luna.
I also have a claude max plan for coding but like I mentioned in the article coding is still separate from the agents.
richardkam0511 | an hour ago
[OP] carimura | an hour ago
aliasxneo | an hour ago
[OP] carimura | an hour ago
avereveard | an hour ago
tln | an hour ago
You haven't put this on github have you?
world2vec | an hour ago