In this article
A successful agent task does not automatically become a repeatable capability. Another team may lack the method, the context, or the execution boundaries needed to reproduce it.
brazosd provides different forms for different capabilities: skills, tools, delegated agents, and cloud applications. Together they let a platform accumulate more than conversation outputs.
Beyond the prompt, a method
A prompt can state today's goal. Ongoing work also contains stable knowledge: which materials to check, which rules to apply, how to handle missing information, and how to evaluate results.
Skills organize that knowledge into instructions and supporting resources an agent can use. Methods can be reviewed, revised, and shared rather than retyped each time.
A skill still depends on interpretation and execution. An instruction saying “always check” is not a system-enforced rule. Permissions and critical business constraints need enforcement in tools, APIs, or application logic.
Different forms serve different purposes
| Form | Question | Role |
|---|---|---|
| Prompt | What is needed now? | State goals and context |
| Skill | How should this kind of work be done? | Preserve methods and acceptance criteria |
| Tool or API | Which explicit actions are available? | Read, query, change resources |
| Agent delegation | Who handles each part? | Divide reasoning, execution, and review |
| Cloud application | Which rules should run repeatedly? | Execute stable logic and preserve state |
Swipe to see the whole table
A research skill can use file tools, delegate different questions, and eventually turn repeatable statistics into a cloud application. The choice depends on what is stable and what requires judgment.
Delegation needs clear boundaries
Several chat windows do not automatically form a team. Useful delegation needs goals, inputs, responsibility, and an explicit result.
brazosd supports sub-agent sessions and waiting for a specified run's final outcome. The target must be accessible to the current user; an agent allowlist can further restrict delegation.
Separate conversations do not imply separate external data. Two agents may still use the same Cloud Drive or database. A sandbox snapshot is not a complete copy of external business state. Concurrent changes and ownership of results still need consideration.
Stable methods can become applications
When rules are well understood and frequently repeated, reinterpreting every step through a model may be unnecessary. An agent can encode those rules in a cloud function and keep business state in SQLite.
This creates a progression: explore through conversation, preserve the method as a skill, and implement suitable repeatable parts as an application. Not every task should reach the final stage. Changing or judgment-heavy work may remain better suited to agents.
Methods, execution interfaces, and services can evolve independently because each has a clear role.
Distribution involves more than copying instructions
The platform already provides ways to discover, use, and share agents and skills. Reusable methods can reduce duplicated workflow design and help turn specialized experience into a product.
Sharing a skill does not grant the recipient its author's files, devices, or enterprise permissions. Users need their own authorized resources, and must still evaluate the source, quality, and applicability of the method.
An ecosystem depends on reliability, stable interfaces, understandable version changes, and clear identity and data boundaries. A marketplace is an entrance, not a substitute for those foundations.
A workspace that retains capabilities
A successful task can leave deliverable files, a reusable method, a tested combination of tools, or a running application. Agents organize the work; the platform provides persistence, authorization, and execution.
Serverless and brazosd on brazosd explain how those capabilities become applications and product features. The skill development documentation covers their practical organization.