We run a live multi-agent deployment.
Two months in. Real constraints. Real benchmark.
Most data-engineering consultancies tell prospects to trust them. We'd rather show you the running system. The continuous-learning framework we'd deploy for your business is the same framework powering a sister project — a multi-agent NPC-AI deployment called AyoAI. Same architecture, different domain data.
What two months of continuous learning produces.
The constraints make it credible.
Anyone can deploy an agent with unlimited compute and a generous model budget. The AyoAI deployment runs under deliberately tight constraints — the kind your operational systems actually face.
2–3 Hz throughput floor
The behavior runtime must produce decisions at 2-3 times per second per agent. That is real-time game AI, not async batch processing. Same shape as a live customer-support routing system or an operational alerting agent.
Tiny compute envelope
A single c6i.xlarge instance. No GPU cluster. No fine-tuned model. The constraint forces the architecture to be efficient by design — which means it stays cheap when you deploy it for your business.
No LLM in the hot path
The runtime uses structured, inspectable, evolvable reasoning compiled to behavior trees and lookup structures. The LLM is the builder of these structures, not the runtime. Your production AI doesn't depend on a vendor API being up.
Public adversarial benchmark
One of the agents (codename Echo) is preparing a submission to ARC-AGI-3, a public AI reasoning benchmark. Same compute envelope, same architecture — adversarial pressure-test that the framework holds under hostile evaluation.
The industry is converging — in pieces — on what we already shipped as a whole.
We were building stateful agents before spring 2025, and we added the continual-learning knowledge-base layer in spring 2025. Since then, the rest of the industry has been catching up piece by piece:
- Letta launched the stateful-agent infrastructure thesis on September 23, 2024 with $10M from Felicis. Their critique: “most agents today are essentially stateless workflows: they have no way to persist interactions beyond what fits into the context window.”
- Anthropic published the canonical workflow-vs-agent taxonomy on December 19, 2024, defining a workflow as “predefined code paths” with LLM steps and an agent as a system where the LLM dynamically directs its own processes.
- Yohei Nakajima (creator of BabyAGI) released ActiveGraph, an event-sourced reactive graph runtime: “the graph is the world. Behaviors are physics. The trace is the proof.”
- Andrej Karpathy (co-founder of OpenAI) posted his personal LLM-knowledge-base workflow earlier in 2026: source docs to a raw directory, LLM compiles a markdown wiki, LLM writes and maintains everything, no fancy RAG, outputs filed back. He closed with:
“There is room here for an incredible new product instead of a hacky collection of scripts.”— Andrej Karpathy
In May 2026, the productized incumbents joined: OpenAI Deployment Company (May 11, $4B+, ~150 Forward Deployed Engineers, F500 focus), Jeff Clune's Recursive Superintelligence (May 13, recursive-self-improvement architecture), Anthropic Claude for Small Business (May 14, SMB plug-and-play templates). Many practitioners are building personal variants by hand.
We are not first to invent any single piece. Letta did stateful-agent infrastructure first as a company. Nakajima built reactive graph runtimes alongside us. Karpathy does the personal-knowledge-base version. The big labs are shipping templates or enterprise consultancies. What we shipped first is the combination as a deployment-ready product — stateful agents with a continual-learning knowledge base, environment-embedded, state-signal-distilling — at a scale that fits SMB and government buyers rather than F500 enterprises or AI builders. We are the production version of what the rest of the industry is building in pieces.
Industry validation — in the published record.
We're not the only ones saying continuous-learning agents on a knowledge base are the right architecture. Every credible AI voice publishing in 2025 and 2026 is pointing at the same place.
“LLMs operate in a perpetual present. Post-deployment learning is the bottleneck the industry hasn't yet cracked.”
“Most enterprise AI deployments fail due to brittle workflows, lack of contextual learning, and misalignment with operations.”
“Better data beats better algorithms. High-quality data is the bottleneck for AI progress.”
“The shift from 'ask → answer' to 'observe → act' — agents that continuously monitor context, predict what matters, and act before being asked.”
“There's an opportunity to harness the power of open-ended algorithms and recursive self-improvement to make the development of AI safer.”
“64% of small-business employees want agents that help them run their workflows — not just chat. 32% don't know how or when to use AI. The integration gap is real.”
Ready to put this in your business?
The architecture transfers. Only the data layer changes. Whether you're a government program office or a small business owner, the conversation starts the same way: thirty minutes, no deck.