The Hallway Track
9 tracked signals on context-engineering.
Microsoft introduces Microsoft IQ to give agents shared enterprise context, addressing why agentic projects fail.
“More than 40% of agentic AI projects are expected to fail. And that's not because the models aren't capable, but it's because the agents lack the context that they need to be able to succeed.”
Anthropic defines agents by leaning on model intelligence with minimal core primitives in a loop.
“Agents in anthropic are actually very simply defined where essentially you want to give the model and lean in on model intelligence as much as possible.”
Engram, a new lab from Dan Biderman and Jessy Lin, treats memory and continual learning as baking new context directly into model weights.
“We don't see the world through the lens of pre-training or post-training. Our models are always training.”
Harness engineering alone can dramatically improve agent performance without changing the underlying model.
“we moved from 30th to 5th on Terminal Bench just by doing some harness engineering, without even changing the underlying model”
The bottleneck for reliable coding agents is not intelligence but specifying, managing, and verifying their work.
“today's agents are mismanaged geniuses”
Databricks aims to make enterprise AI work by feeding organizational data, processes, and tacit knowledge as context to frontier and open-source models.
“If we just give the context to these very smart models that can solve those super hard problems, they'll be able to do amazing things inside of our companies.”
More context degrades agent performance because LLMs ignore the middle of long contexts in a U-curve pattern.
“This is like a U curve where some of the things from the start, some of the things from the end make sense but whatever you are providing in between that that is not taken up.”
Neo4j proposes graph representations to give AI agents richer context from lakehouses than queries alone
“context coming in shapes and not necessarily queries”
AWS shows how to build context-isolated research agents using LangChain Deep Agents and Bedrock AgentCore subagents.
“A better approach is to delegate deep work to isolated subagents that return only concise results.”