The bottleneck isn't your AI model — it's that your institutional knowledge isn't built for agents to consume

A technically skilled engineer joining an enterprise faces a structural barrier that no amount of talent compensates for: they lack the institutional context that makes decisions here make sense. Why does the payment reconciliation service split transactions across two ledgers? Because a regulatory change in 2019 required separate audit trails for domestic and cross-border payments. No amount of reading the codebase will surface that fact — it exists only in the heads of engineers who were there.

Bakal calls this the Institutional Impedance Mismatch: the structural disconnect between a knowledge consumer's existing knowledge and the institutional context needed to act correctly. The same mismatch that afflicts new engineers afflicts AI agents — and at every scale. Modern coding agents like those evaluated on SWE-bench perform well on open-source repositories precisely because open-source knowledge is encoded in forms they can access: the repository itself, the contribution guidelines, the public APIs. Enterprise software development operates under fundamentally different conditions. Decades of institutional knowledge — architectural decisions and their rationales, deployment procedures tailored to specific infrastructure, compliance policies mandated by regulatory frameworks — exist as tribal knowledge, distributed across wikis, chat threads, runbooks, and the heads of experienced staff.

Retrieval-Augmented Generation (RAG) attempts to bridge this gap by surfacing relevant documents at query time. But RAG retrieves — it does not curate. It surfaces existing documents without verifying their relevance, currency, or completeness. When the knowledge exists only as undocumented tribal knowledge, RAG has nothing to retrieve.

Bakal's proposed solution is Knowledge Activation — a framework that builds on the AI Skills open standard for agent-consumable knowledge. AI Skills provide a structured format for encoding knowledge that agents can consume directly, without the interpretation overhead of prose documents. Knowledge Activation specializes Skills into Atomic Knowledge Units (AKUs): tightly scoped, governance-aware knowledge artifacts designed specifically for institutional knowledge delivery in enterprise software contexts.

What distinguishes an AKU from a retrieved document is its action-readiness. Rather than providing background information that an agent must interpret to determine what to do, an AKU encodes exactly what to do, which tools to invoke, what constraints apply, and which related AKUs to consult next. Each AKU is aware of the others: it declares typed relationships — dependencies, successors, related systems — that form a traversable knowledge graph at runtime. An agent navigating an unfamiliar codebase doesn't need to interpret a 40-page architecture document; it can traverse the knowledge graph, reading only the AKUs relevant to its current task, in the order they're needed.

The paper formalizes the resource constraints that make this architecture necessary and specifies the full AKU schema and deployment architecture. It also grounds long-term maintenance in knowledge commons practice — the idea that AKUs, like open-source code, must be governed, versioned, and maintained by the teams that benefit from them, not centralized into a documentation team that quickly becomes a bottleneck.

An AKU is a structured markdown file with YAML frontmatter encoding its metadata and relationships:

The agent reads this AKU, knows immediately which tools to call, which constraints to respect, which compliance requirements apply, and which AKU to consult next. No interpretation required — just execution.