Context is the
bottleneck

The information exists. It's in the repos, the tickets, the docs, the conversations, the logs; most of all it's evident in the system behaviour. However, as we've all personally seen it's contradictory, stale, unlinked, and sometimes hallucinated. Agents and humans are reasoning over the same mess.

Context defines the token search space meaning that an agent's ceiling is defined by information completeness and density. We've seen successive waves of innovation in the space from RAG and other code-search tools, to memory systems that recall past conversations with embedded lessons, and all the way to knowledge graphs.

But RAG retrieves by similarity, not by the true causal relevance to the task. Memory systems store agent thoughts without verifying them. Knowledge graphs often formalize the contradictions via inconsistent ontologies - and are rarely time-aware or causal. So agents reason over contradictions and confidently produce wrong answers.

Iterance started out of the idea that the truth about every software firm is embedded in its tech stack and would be valuable to investors. We quickly realized the real value isn't the agents for broad assessment - it's making the relevant verified facts about the firm efficiently retrievable by all stakeholders. Every fact checked against related facts. Contradictions resolved, not stored. Causal structure connecting code to architecture to product concepts to outcomes.

That's what we're building. A symbolic representation of organizations that lives outside the LLM - a verified, non-conflicting knowledge base that agents and humans can reason over and edit with confidence. When the context is clean, our agents thrive.

We're early in our development but are following several convictions in our research. Context should self-evolve and agents should participate in the change-making. They should continuously verify, deduplicate, and extend the knowledge base. As a firm and its tech stack changes, the knowledge base keeps up.

We think context verification infrastructure will become as fundamental to software teams as CI/CD is today; a layer that every agent and every engineer utilizes to generate and maintain the current source of truth.