The project
Files, branches, tests, and repository instructions describe the current application and the rules for changing it.
LEARN · CODING AGENTS
What does a persistent coding agent keep?
A persistent coding agent works in a project that can continue across tasks. The useful continuity comes from retained files, repository state, task records, and conversation history that can be inspected and revisited.
Three kinds of continuity to understand.
The work record
Task scope, conversation history, tool activity, and review results explain what was attempted and what remains unresolved.
The next task
A follow-up task can use retained project state, while checking the current code and clarifying any changed requirements.
Persistence is not unlimited model memory.
A model processes a bounded context for each request. Keeping a workspace and a conversation history does not mean every detail is included in every future model call. Useful continuity depends on retaining the right artifacts and selecting relevant context when work resumes.
Repository instructions, a concise task description, and recorded check results help a later session reconstruct the work. When the code or requirements have changed, the agent still needs to inspect the current state rather than assume an earlier conclusion remains true.
Worked example: continue improving a support portal.
The first task adds customer tickets and checks that each customer sees only their own records. The result includes the code changes, tests, and a review of the access behavior. The next task asks for a filter showing unresolved tickets.
A continuing agent can inspect the existing ticket fields and access checks, add the filter within that design, and rerun the relevant checks. The earlier work provides context, but the new change still needs evidence that customer isolation works with the filter applied.
What a useful task handoff contains.
- The intended outcome and the behavior that must stay the same.
- The repository, branch, and relevant files or documentation.
- The checks that ran, their results, and any remaining uncertainty.
- The decisions or access needed to continue.
- A clear distinction between a proposed change, reviewed code, and a verified deployment.
How this fits Redgold.
Redgold provides coding-agent workspaces and surfaces for conversations, tool activity, and code review. Its managed application workflow connects reviewed application artifacts to routing, execution, data, and deployment.
Permissions and the configured workflow determine which actions an agent may take. Persistent work also needs lifecycle management: a retained project does not guarantee that an execution process runs continuously or that an interrupted task completes without intervention.