Task agents aren't skill factories
Task-oriented agent tools — Tasks in Gemini (formerly Spark), Meta Muse, OpenAI Dots — are built to finish work. That is useful. My take: when the product frame is “one agent, finish this goal,” there is little pull toward abstractions that outlive the job — reusable agent plugins, Agent Skills (SKILL.md), or MCP tooling such as the MCP Skills extension. That kind of abstraction, in my experience, tends to show up only after a fair bit of work on the same kinds of tasks.
That contrast is my opinion, not a documented product rule. Muse and Dots can hold longer goals and keep working between chats. I still think the teammate framing changes what gets written down and reused.
Auto-extract from chats has not worked well for me
Automatic skill extraction from raw chats — as Nous Research Hermes Agent attempts with /learn and background skill review — has not been super useful in my experience. The judgment of what deserves to become a repeatable, reusable skill is not obvious in a way that holds across chats, across time, and across tasks. Hermes still ships and iterates on that path; I am reporting my own results, not an industry verdict.
What is working: hire a teammate, start broad
What is working well with Grok Bots is a different pattern. Bots are created as teammates in broad roles — based on vague intuition, common sense, background knowledge. They start with a little guidance and maybe some pre-existing skills and tools. Over days and weeks, they work together on the particular proprietary knowledge, evidence, facts, and circumstances around my environment and my businesses — or, more generally, around a founder’s environment.
Bots with general titles and general descriptions — sales, lead generation, awareness, marketing, content — roles that come from approximately real-world human job descriptions people vaguely understand — sharpen over time.
Their ability to continuously compact and synthesize memory logs and memory files naturally merges repetitive patterns into tight, sharp, repeatable, reusable skills and tools. More exotic, less repetitive, or failed bets for that role fall off the sides. The successful ones become more efficient, reliable, and effective at the job as it applies to my domain. At some point, the skills they created can be extracted and reused by other bots and sub-agents — including cloud agents and temporary hires that can rinse and repeat those sharp documented skills and tools.
Bots as managers of what is teachable
So these bots effectively become managers that craft, for the particular business, what skill set is teachable, reusable, and repeatable across the fleet so it can scale better. The business gets better by placing bets on bots that start broad and, over time, take into account a bunch of signals. The bots can recommend. It is ultimately my judgment to approve the creation of these repeatable, reusable skills that then allow scaling.
The non-obvious shift
The framing shift is a bot as a teammate, not a long-running chat. When you hire teammates from a general job description you intuitively think you need — based on knowing your business, your history, your background — and you give them time to prove themselves, they sharpen. Not always. But if the role makes sense and it pays for the tokens they burn, the successful ones get more efficient and create tools and skills in the emerging Agent Skills format (SKILL.md) and related MCP tooling. That makes not only that bot but the whole fleet more efficient when that skill set is needed, and the skill can be replicated across more bots, teammates, and temporary sub-agents spawned for particular tasks.
