From Interactive Work to Continuous Engineering

The shift

An interactive coding Run starts with a human request from the workspace-root driver. A housekeeper, dependency maintainer, or bug fixer starts with a schedule, trigger, or queue owned by an external orchestrator. The execution model can remain the same: each role supplies caller-owned Task intent and invokes a bounded ws run against an isolated managed-repository worktree.

schedule / trigger
        ↓
bounded Run → inspect → propose → validate → checkpoint
        ↑                                      │
        └──────────── next Run reads ─────────┘

Continuity comes from durable Workstream goals, loop definitions, state, history, and semantic checkpoints owned by the caller. It does not require keeping the same process, Task Workspace, worktree, or agent session alive forever. AGENTS.md and .agents/skills/ are regenerated when the System Workspace map or shared conventions change.

Roles, not personalities

An Agent Profile describes an engineering role: interactive coder, housekeeper, dependency maintainer, or bug fixer. The profile can define defaults for tools, scope, and validation without becoming a mandatory implementation policy.

Safety boundary

Scheduled Runs should produce Candidates and evidence. They should not silently bypass repository review, resource limits, or integration policy. Scheduled Housekeeper describes the default.

The bigger picture

Continuous engineering composes an external Planner/orchestrator with bounded ws primitives: context composition, isolated Task Workspace execution, explicit coordination, result/lineage records, durable knowledge promotion, and separate integration.