Over the past year, every major AI coding harness has settled on the same way to learn a task or internalize a skill. A skill.md file describes one procedure, from writing a database migration to cutting a release.
While skill files have filled an important gap left by today’s rudimentary agent memory, especially in a multiplayer environment, they can quickly become disjointed, stale, conflicting, and harmful to valuable agentic output.
Skills are now evolving into knowledge operating systems that are highly tuned to the business, far more maintainable, built for multiplayer, and thus far more valuable to the business.
Skills often overdrive newer frontier models
Generic skills dominated AI software engineering spaces in early 2026. However, teams soon realized that generic skills for writing tests, understanding frontend development, or reviewing pull requests end up explaining things that future frontier models will be far more skilled at delivering.
In fact, by mid-2026, the engineers behind Claude Code were advising users to scrap their skill files regularly. At YC's Startup School in July, shortly after the launch of Opus 5, Boris Cherny, who created Claude Code, told founders to delete their CLAUDE.md, skills, and hooks every six months because "the model might not need all those instructions."
Knowledge operating systems are much more delicate in their approach, leaning on the inherent evolution of frontier models and superintelligence rather than purely deterministic guidance. They form a basis for understanding what makes work valuable instead of specifying exactly how work should be done.
Your operating system should guide with vision and allow probabilistic power to flourish
Rather than overdriving models with explicit instructions, agentic guidance is now more about management, specifically the management of something far more intelligent than the vast majority of humans and engineers.
But managing something that is vastly more intelligent than you, or someone who is, has always been a delicate process. In this scenario, as in the past, your job is to nurture the extreme ability of a worker and never dampen it, while also guiding and focusing that ability into valuable output.
To achieve this in an agentic setting, the best knowledge operating systems focus on capturing tacit knowledge and processes, particularly what "good" means to a team. Models typically have a wealth of explicit knowledge, as math, science, and engineering operate universally, but they'll fill in the gaps on what you expect them to achieve relative to what your team values. That's what your team needs to define in the operating system.
This knowledge capture extends not only to ways of working but also to ways of thinking. For example, models tend to be highly risk-averse, which means vastly more valuable outcomes can be achieved when you continuously reinforce ambition along the path of a greater vision.
However, models also tend to be highly goal-oriented, which means a model will apply brute force and tremendous effort to a path, whether that path is the best one or not. So it becomes your team’s job to temper your agents’ goal-seeking with vision, higher-level direction, and the ways of working your team has used to produce valuable work in the past.
The most efficient way to achieve this outcome is by capturing your team's tacit knowledge. Imagine a talented new team member joining your company. How would you set them up for success? What would you tell them?
The end result is a workflow that makes your team’s tacit knowledge obvious not only to your full team but to your agents as well. This knowledge becomes a powerful shared company brain that people and agents can use together as thought partners.
In the end, you'll have a finely tuned knowledge operating system that allows the probabilistic power of any model to flourish while also building value in the ways your team implicitly expects.
Knowledge operating system process and approach
There are specific types of captured knowledge in a knowledge operating system that build more value than others. These can be organized into an easily navigable structure that may include:
- Routing and context scope: Simple, direct routing at the root level, beginning in AGENTS.md, allows agents to quickly traverse to exactly the context they need and nothing more. Codespeed Context Profiles further enhance this focus with unique specialties per agent or Thread.
- Specifications process: Technical specifications and requirements are an important part of agentic software engineering. A standardized process for drafting those plans, then having additional agents adversarially review and verify them, is essential.
- A shared company brain: Distilling all agentic and human thought into a single source of truth ensures that everyone on the team is working from accurate, useful knowledge. Your operating system then becomes a shared company brain that agents can easily parse, with structured folders, `.md` file entries, and processes for creating or updating entries. This format is also highly valuable for humans who use agents as conduits and thought partners. Together, agents and people keep the shared company brain up to date and focused. Codespeed includes specific knowledge tools to automate this process so you can focus on the most valuable work at hand.
These are just a few of the types of knowledge structures that can exist in your operating system, each highly tuned to the unique needs of your team.
Codespeed knowledge automation tools also ensure your knowledge operating system acts as a very organic source of truth rather than a continuously patched set of thoughts and memories that require constant maintenance, including tracking knowledge freshness, drift, and accuracy.
Codespeed supports both the Codespeed Open Knowledge Protocol, released on March 11, 2026, and Google's Open Knowledge Format, released on June 12, 2026.
