A change request starts as a short message. Then come documents, questions, versions, another team’s dependency, and approval the following day. Giving AI agents responsibility for that work requires designing what happens between moments of execution.
The assignment needs its own working record
In our approach, an ongoing process needs a state that can be revisited: the original objective, work plan, accepted outputs, approved decisions, and remaining blockers. That record allows the next agent to understand its assignment without reconstructing everything from a long conversation.
Consider a software project. A review discovers an unhandled error path. The working record needs to connect the finding to the requirement and change version, return responsibility to the appropriate role, and identify what must be checked again. A generic message saying it needs fixing leaves too much reconstruction to the recipient.
Useful memory preserves distinctions
We distinguish source material, task state, and knowledge that may be useful later. A received document is a source. A decision to wait for approval is working state. A missing review rule may become an improvement proposal. Each has a different lifetime, owner, and permission boundary.
This distinction also affects quality. An early draft should not become an approved decision merely because it was saved. An assumption valid for one client does not automatically become a company rule. Preserved information needs context explaining when and why it may be used.
Anthropic’s context-engineering article describes compressed history, persistent notes outside the context window, and specialist subagents as approaches to extended work. Our design inference is to make continuity explicit and deliberately choose which information returns to context at each stage. Effective context engineering — Anthropic
Depth requires a handoff agreement
An extended process can branch into several subprocesses. One manager owns the objective, a subordinate manager coordinates a workstream, and specialists perform focused checks. Each assignment needs to specify what comes back: an output, supporting evidence, limitations, and an interpretable status.
In AOS, the management structure supports that continuity. The AI CEO prioritizes and delegates; agent managers coordinate dependencies and review results. Returned work carries a reason and its context. When an assignment reopens, previously accepted work remains part of its record.
An interruption is a state to design for
Sometimes a source is unavailable. Sometimes a person must decide, or the assignment is no longer relevant. We suggest defining a valid pause, what must be preserved, and who becomes responsible for continuation. This distinguishes work waiting for a known reason from work that has lost direction.
Resumption needs similar care. Before repeating an action, the system should establish whether it already happened, whether approval was received, and whether the underlying information remains valid. For draft preparation, this can prevent duplication. For a process connected to external actions, it becomes a basic execution design requirement.
Evaluate continuity as well as the impressive moment
When evaluating a process, ask to see a changed requirement, a conflicting source, feedback that sends work backward, or a break between working days. Does the agent identify the change? Does it retain what is still valid? Can the person understand what is needed from them?
For us, depth means holding an objective across that sequence. A discussion of models, prompts, and tools therefore includes memory, state, delegation, and review. These components support the move from an assignment that succeeds once to a process that can be operated, understood, and improved.
Ask how the process pauses, gets corrected, and resumes. That reveals whether the foundation for extended work exists.
Explore AI teams in action
