The Hallway Track
18 tracked signals on agent-architecture.
NVIDIA's AVO agent architecture achieves 100% on ARC-AGI-3 benchmark.
“A frontier language model is only one component of an AI agent. The surrounding agent system—often called a harness—determines how the model receives context, uses tools, maintains state, responds to feedback, recovers from failure, and sustains progress over long-running tasks.”
Screener sub-agents read full traces so the main agent's context stays manageable
“we're really familiar with designing an org chart, in a way”
Navan architects argue AI agents are at the same inflection point as microservices in 2015
“If you can't build a single agentic loop, why go in and try to build a multi-agent orchestrated system?”
Satya Nadella frames agent architecture as a harness looping across models, data, and tools.
“you kind of want to harness to define the models, the the data, uh and the tools. And so that you have a loop across those three.”
Harmonic achieved 4x user retention by replacing a complex LangGraph pipeline with a simple Deep Agents tool loop.
“I think everyone's secret weapon is becoming Scout.”
BabyAGI creator proposes inverting agent architecture to build around the event log, not the LLM
“ActiveGraph asks, what if you build around the log?”
Agent architectures decay in 6 months; design for layer-based durability to survive AI churn.
“if you've been building agents for more than 6 months, you've likely rewritten something, maybe more than once.”
Warp uses outer loop agents to automatically improve inner loop agent skills over time
“you have another outer loop agent that watches the inner loop agent's work and improves its skills over time by detecting errors or taking into account user feedback”
Enterprise AI agents waste effort rediscovering data sources; a shared ontology-based semantic layer fixes this.
“every single time a team has to build an agent, they have to figure out from scratch where the data that they require for that agent to operate, where it sits”
LLM performance degrades after 25% context window usage, making 'infinite context' a dangerous myth
“if we consider that there's infinite context window, we just think that we can dump the whole list of documents to the context and expect it to magically work”
Human-focused API architecture fails at agent scale, requiring ground-up rethinking of auth systems.
“the human-focused architecture doesn't scale well for agents”
An agent's true identity is its append-only log, not the model or runtime executing it.
“The agent is its data. It's specifically the log.”
MCP's core value is isolating API auth flows outside the agent's context window, not tool definitions.
“Maybe the idealized form of MCP is just an auth gateway for the API and nothing else. That'd still be a win.”
Supply-side agents capture org context and decisions without user access by living where work happens.
“you don't really need to have access to the user. What you need access to is the system of context as well as decisions are being made throughout the organization.”
Google recommends five architectural patterns for production-ready AI agents at scale
“Production agents need robust architecture, not just clever prompts.”
LangChain's managed deep agents support on-demand skill loading via progressive disclosure
“Skills are basically context and scripts that your agent can run and read dynamically as it's running.”
A vendor combines AI agent swarms with a high-scale data platform to prioritize enterprise security vulnerabilities.
“agents are actually coloring the graph over time to create more and more interesting lower high confidence edges”
A walkthrough explains the Hermes AI agent's architecture: agentic loop, memory, context, and messaging gateways.
“the gateway is the system that is always running that allows you to connect to your um Hermes agent via Telegram, email, Slack, etc.”