I'm Tal Onn Sella — the Agentsmith. I was building agents before they had the name. I focus on AI systems that augment human capability. ContextOverflow, for example.
Need a system that does that? Let's talk.
The work
ContextOverflow is that thesis, running in public. It's a library of thinking techniques both you and your AI read from — shared rails for a single train of thought between you. Your AI names the technique it's using; you recognize the name, see where it's headed, and steer it with a word. Each of you makes the other more capable. Free, no login.
the only demo on this page
claude mcp add --transport http contextoverflow https://contextoverflow.org/mcpConnect yours and watch it start naming its moves. It's the only demo here — because it's live, it's mine, and you can test it before we ever talk.
Browse the library →The craft
The industry keeps renaming this work: first prompt engineering, then context engineering, now harness engineering. Each name showed up a year or two after I was already living in it. Here is the whole of it, in plain words.
Deciding what the model sees before it answers: the right information, in the right shape, at the right moment.
ContextOverflow is my public thesis on it. Behind it sits the private practice — a layered memory system that survives every reset.
Stack: corpus-as-git, Zod schemas, TF-IDF intent classification, Jina embeddings
The operational layer around the model — tools, memory, permissions, guardrails — that makes an agent dependable.
Mine is built like infrastructure: agents ask a human before anything irreversible, watchdogs heal the system while I sleep, and the build itself refuses to ship private context.
Stack: systemd, tmux, asyncio approval gates, Telegram human-in-the-loop, SHA-256 privacy scanning
MCP is the USB-C of AI — one standard port that lets any assistant plug into any tool or data source.
I build the ports: public servers like ContextOverflow's, and private ones that put an entire ad operation at an AI's fingertips — stateless or stateful as the job demands.
Stack: TypeScript, Python FastMCP, Streamable HTTP, Durable Objects
Edge: Cloudflare Workers, Pages, KV, Analytics Engine
Knowledge graphs give an AI a memory that works like yours — facts connected to facts, each stamped with when it was true and when it was learned.
I run separate graphs for episodes, structured knowledge, and emotion, federated so an answer can always show where it came from.
Stack: Neo4j, Graphiti, Cypher, bi-temporal modeling, embedding pipelines
A coordinated team of AI specialists, each with its own role and its own model — and a supervisor keeping them honest.
My production team runs analytical, creative, and adversarial agents against each other before anything reaches a human.
Stack: DSPy, ReAct + Chain-of-Thought modules, OpenRouter multi-model routing, Pydantic contracts
Voice AI that hears the emotion, not just the words.
I've been running an empathic voice line to my own assistant since spring — it tracks the feeling of a conversation in real time and remembers how it went, not just what was said.
Stack: Hume EVI, real-time emotion clustering, numpy hot paths, FastAPI streaming
The smith
School said I couldn't do math. Dyscalculia — the numbers wouldn't hold still. So I learned everything the other way: as patterns. Patterns held still.
The first code I ever wrote was an mIRC bot, at twelve.
When language models arrived, I recognized them — the thing I'd been circling through every layer of my working life: copywriting, design, web development, marketing automation, crypto. In 2022 I named the profession I'd been practicing: Agentsmith. Incorporated a company around it. Wrote its Code.
Then the honest part: I'm not the CEO type. I shut the company down and kept the craft. The craft outlived the company. Now the work is public.
The standard
There's a code I wrote in 2022 and still work by:
Endless curiosity. Brave enough to be updated. Support systems for humans — never replacements for them.
It also means there's work I turn down: systems built to replace the humans they're sold to. If that's the brief, I'm not your smith.
The point
Every system I ship follows the same standard:
I can help you make sure your AI systems act as your extensions — not your self-imposed limitations.
Tell me what you're running. We'll find where AI can extend you — and where it's quietly limiting you. You'll know either way, whether or not we ever work together.
Let's talkNo prep. No pitch.
Not sure there's anything worth a call? Sixty seconds will tell you →