Software Architect | AI | .net | DevOps | k8s | Foudner cairnriver.io
London
- Tweets 127
- Following 738
- Followers 288
- Likes 543
Pinned Tweet
i'm Spencer, 50, from London co-founded and sold a few software businesses, now trying to solo launch a new one.
keen to connect with anyone building cool stuff in AI or indie game devs.
love getting into the nerdy low-level details 🤓
also anyone who is going to be playing warcraft forever when it launches (productivity about to take a critical hit)
I'm 29.
Solo founder from Brazil, based in Barcelona.
Looking to connect with more marketers & indie hackers!
35
29
1,493
When playing around with Local LLMs or using them in agents, it is mostly "single file" traffic. You're the only user of the model and until you start running multiple agents at once or agents that fire off parallel sub-agents - the overheads of running multiple requests aren't an issue.
In this article I dive into what happens when you try to push an already limited 8GB GPU to run multiple requests at once, seeing what works and where it breaks.
Feedback/criticism of method or results welcome!
nitter.net/spencerclarkdev/status…
4
78
nitter.net/spencerclarkdev/status…
Jumping on the Jev-train and added support for it to Vessel - now you can visualise your inputs and outputs in the proxy
Vessel 0.2.0: replay one prompt across several models, or one model across a temperature sweep, score the responses 1–5, or ask your agent to judge via MCP. Also: named sessions via headers, reports, CSV/JSONL export, brew and scoop installs. vesselproxy.app
1
5
115
Toadstool Tactics - a silly little game i've been working on for fun.
fatpunks.itch.io/toadstool-t… free to play on itch.io
1
3
96
I'm building with Builders of X 🚀
Find builders from around the world:
buildersofx.com/builder/spen…
1
70
I think this is great advice. Providing all the ugly domain specific plumbing your agents need to succeed is key, and not something generic harness can support as easily.
Learn to build a harness, folks.
It's not surprising to me that so many YC builders want to build domain-specific harnesses.
If you work long enough on a domain-specific problem, you quickly realize the opportunity. But you also realize how important that harness will be to stay competitive in the agentic era.
From a product perspective, harnesses open up interesting new surface areas and experiences for the services/products you provide.
From a technical perspective, harnesses are how you build and maintain a framework and set of best practices for how your users/customers interact with what you offer.
Understanding how to build and design a harness means you can build much stronger intelligence stacks, given that you can customize it and understand the domain well. That's extremely valuable. It may not seem apparent yet, but a harness wave is coming.
If you are getting started, give this list of harness papers to your agents and start upskilling: academy.dair.ai/papers/colle…
If you are already a builder, try building one for your specific domain. It's a lot of fun, and you learn a lot of interesting things to enhance your current agentic tools.
1
5
281
i'm 50, from London
co-founded and sold a few software businesses, now trying to solo launch a new one.
keen to connect with anyone building cool stuff in AI
love getting into the nerdy low-level details 🤓
2
4
102
I wanted to play with that idea of letting a bunch of agents try and run a D&D style game together.
Give them a room, a dungeon master and a rulebook - implemented by a game engine inside the harness. The engine was the single source of truth, so agents couldn't conjure up whatever they liked.
Would the agents stick to the rules? Would they work together? Would their personalities affect their decisions? Would they even remember whose side they were on?
It also looks at how different models played the game.
2
3
75
Vessel 0.2.0: replay one prompt across several models, or one model across a temperature sweep, score the responses 1–5, or ask your agent to judge via MCP. Also: named sessions via headers, reports, CSV/JSONL export, brew and scoop installs. vesselproxy.app
2
3
192
Just dipping my toes into the world of Python AI, coming from the .NET world. It seems like LangChain/LangGraph is the dominant force over here.
It took me a minute to get my head around why it was needed - hiding away a lot of boilerplate - but I guess I’m just used to rolling up my sleeves and building that framework layer myself.
While digging, I’ve noticed a few alternative ecosystems: CrewAI, PydanticAI, LlamaIndex, Model-native SDKs (OpenAI/Anthropic)
What are people's thoughts on these? Are they up-and-coming and worth investigating alongside the LangChain stack, or is LangChain still the clear winner for production?
4
1
3
96
Bit of fun with GPT Astra - a one-shot to create a physics heavy 2D platformer in Unity - one prompt, 52 mins then a smaller 2nd prompt to tidy up some animations. On High.
1
12
4,390
workflow question: I had been using "claude chat" to discuss ideas, changes, bugs. Then when happy I'd ask for a prompt for me to give claude code to do the actual implementing.
For the last week I've changed to doing most of my discussion in claude code, getting it to build out full specs and phase plans MD files.
Can't say really which has been better, other than CC seems more likely to get on with it. Same model(s) but chat agent vs code agent.. which do you use? Or some other way?
1
75
At the start of the year I was very much - "I want to write my own code still, not use AI" - as the year went on it changed to "ok, i'll let it write the tests; fine, it can do this refactor" through "I've done the fun bit, AI can pick up the pattern and implement the rest now its boilerplate", all the way to "claude - here's the design, give me a plan and start step one".
Not quite sure how I feel about that. Getting more stuff done feels good, but still feels like a hollow win.
Consolation is that at least my experience still has a place in making sure the big picture plan is done properly.
1
66
1/ I wrote a small observability proxy that records everything my LLM clients send. Point base_url at it, get capture, full search, token/latency metrics, and replay — all local. Allows different clients/agents to send their own tag to make filtering by agent easier.
1
1
79
2/ Replay is one of the cool features I added: re-send any captured request to a different model and compare responses + tok/s side by side. Also has a read-only MCP server, so your coding agents can query your own traffic logs realtime - pretty useful for debugging work!
1
1
51
3/ Open - MIT, single binary or docker compose. Works with Ollama, LM Studio, llama.cpp, vLLM, OpenAI, Anthropic. vesselproxy.app
8