AI agents may be processing hundreds of lines of instructions to complete tasks that require only a handful of details. According to Modus, median context-loading precision is below 1%, while even results at the 90th percentile remain below 10%. The finding highlights a growing challenge for enterprise AI: providing agents with relevant, up-to-date knowledge without overwhelming them with unnecessary information.

In a recent blog post, Tomer Mesika, Co-Founder and CTO of Modus, argues that static AI skills are contributing to this problem. While these instructions encode organizational workflows and technical knowledge, they also contain operational details that inevitably change. As companies deploy more agents, keeping these instructions accurate becomes increasingly difficult.

Mesika proposes moving toward dynamic skills, where information is assembled at runtime based on the task, the current state of the organization, and the user's permissions.

Why Static Skills Are Becoming a Liability

Static skills often combine two different things: instructions describing how work should be performed and operational facts needed to execute it. The first might include deployment procedures or incident response practices. The second includes database names, cloud account IDs, API endpoints, Slack channels, and feature flags.

The problem is that these facts change constantly. A database gets renamed, a team changes its communication channel, or an infrastructure resource is replaced. The skill remains unchanged unless someone remembers to update it.

Mesika compares this to using a printed map after the surrounding infrastructure has changed. The instructions may look reliable, but their underlying information is no longer accurate.

MCP addresses part of the challenge by allowing agents to retrieve live information from connected systems. However, freshness does not guarantee relevance. MCP is the communication mechanism; the missing piece is determining what information should travel through it for a particular task and user.

The Shift Toward Dynamic Skills

Mesika's proposed solution is to separate organizational processes from the facts required to execute them.

Instead of hard-coding an account ID or communication channel into a skill, organizations can use meaningful placeholders such as "Production AWS account ID" or "On-call channel." The underlying process remains stable, while the corresponding information is resolved when an agent needs it.

"Static skills are essentially caches without invalidation. They preserve a snapshot of what an agent needed to know when someone originally wrote them, even as the underlying organization changes," said Mesika. "The goal should be to keep the process durable while computing the facts an agent needs for each task, based on the latest available information and the user's permissions."

This approach allows agents to receive just-in-time skills containing only the operational details relevant to their immediate requests.

Why a Context Warehouse Is Necessary

Making dynamic skills work requires more than rewriting instructions. Organizations need infrastructure capable of collecting, connecting, verifying, and continuously updating their operational knowledge.

Mesika outlines an approach that pulls information directly from authoritative sources, including cloud accounts, Slack, Git, databases, feature flags, and internal documentation. Rather than treating this information as disconnected text, the system maps entities and their relationships.

For example, an agent working on a payments service should be able to identify its production account, associated database, kill-switch flag, and on-call channel. Each fact also needs metadata covering its source, last verification, ownership, and access permissions.

This is the foundation of a Context Warehouse: a continuously maintained layer of enterprise knowledge that supplies AI systems with the information they need.

How Modus Applies the Approach

Modus is developing this architecture through its platform, using connectors to gather organizational context from existing enterprise tools. Its Context Warehouse organizes information into context items with source details, hierarchical relationships, and permissions inherited from the original systems.

When an agent connects through MCP, it communicates its intent rather than requesting a specific file. Modus selects and assembles the relevant information, filters it according to the user's permissions, and measures how much of the available context was actually needed.

Permissions are also checked with every read rather than relying on previously cached access information.

For enterprise leaders, the takeaway is that improving AI agents may require rethinking the infrastructure behind the instructions they receive. Building dynamic skills involves continuous synchronization, entity mapping, metadata management, and permission-aware retrieval. These capabilities go well beyond prompt engineering or creating more specialized instruction files.

As organizations move toward more autonomous AI systems, the challenge will be giving agents access to knowledge that reflects how the business operates today. Static skills can preserve how work should be done, but a Context Warehouse can provide what agents need to do it now.


This article was published under HackerNoon's Business Blogging program.