Context consistency
How should a change to the campaign brief invalidate or update dependent work?
Technology / Design direction
Orchestration is the work
between the agents.
Our technical focus is how specialist tasks share context, respect permissions and respond to failures. This page describes the proposed architecture and the questions we intend to validate.
Proposed architecture
A campaign is a stateful process.
A good result depends on more than generating an asset. The system must know which brief is current, what has been approved, what changed and what is allowed to happen next.
01 / Shared campaign context
Goals · audience · source material · brand · constraints
02 / Coordination
Task planning & dependencies
Track the work and pass the right context to each specialist.
03 / Specialist tasks
Research · strategy · creation
Prepare outputs against the same brief and review criteria.
04 / Approval boundary
People authorise sensitive and costly actions.
05 / Connected execution
Approved actions only
Proposed integrations must respect tool permissions and campaign limits.
06 / Evidence & review
Results inform the next decision
Preserve observations, recommendations and the reasons for a change.
Engineering priorities
The difficult questions are the useful ones.
Reliable recovery
How should an interrupted task resume without repeating an external action?
Quality before execution
Which checks can be automated, and which need a person with domain judgement?
Decisions under uncertainty
When results are delayed or incomplete, when should the system recommend a change or wait?
A testable proposition
What the approach needs to improve.
| Approach | Useful for | Sadivi’s design question |
|---|---|---|
| Individual AI tools | Preparing an asset or answering a question | Can shared context reduce repeated briefing and inconsistent outputs? |
| Fixed automation | Moving work through a known sequence | Can coordination handle changed briefs and exceptions reliably? |
| Manual delivery | Expert judgement and flexible execution | Can routine coordination be reduced while preserving quality and oversight? |
Product principles
Control has to be designed in.
Action permissions
Separate reading, drafting, recommending and executing. Define which actions require explicit approval.
Traceable decisions
Record the brief version, relevant inputs, decision and approval that led to an action.
Bounded learning
Review changes against the objective and available evidence. Do not treat every movement in a metric as proof of impact.
Next step
Follow the validation plan.
We will assess the architecture against real delivery tasks before making broad claims about autonomous execution or performance.
How we will evaluate it