Research / System architecture
Three agent types.
One connected system.
Communication, activity and production have different responsibilities. Their connections carry the designer’s direction through the work.
A separation of responsibilities
Conversation.
Coordination.
Execution.
For years, Brother Hills has been researching how AI can make production more efficient and cost-effective. This architecture organizes that work around three types of agents, with quality review across the production process.
The interface with people.
Interprets the conversation with designers and communicates questions, progress, issues and completion through the agreed project channels.
- Receives
- Messages, design direction and responses to clarification.
- Passes on
- Request information for coordination; status and deliverables for the people involved.
Focus: what the team needs to know and communicate.
The coordination of the work.
Organizes reception, classification, routing and delivery. Intake establishes what needs to happen; Delivery closes the process with the checked outputs.
- Receives
- The request, requirements and information about the work’s current state.
- Passes on
- A task for the responsible production agent, or a checked delivery for the project channel.
Focus: what happens next, and which role is responsible.
The execution of the task.
Transforms the supplied design into the required assets. Adaptation, Content, Motion and Print each work on a defined production task.
- Receives
- Key artwork, task-specific direction, content and output specifications.
- Passes on
- Produced assets and their requested versions, ready for quality review.
Focus: carrying out the work with the tools the task requires.
Checks outputs against the original request and agreed specifications. Corrections return to the responsible production agent; checked files continue to Delivery.
The shared production context
The brief travels
with the work.
A useful handoff includes the information needed by the next role. The designer’s direction remains the reference as the task moves through production and review.
- Design reference
- Approved key artwork or design system.
- Task & direction
- What needs to change, adapt, animate or prepare.
- Content & versions
- Copy, language and campaign version requirements.
- Output specifications
- Dimensions, formats and production requirements.
- Priority & timing
- The delivery priorities and deadlines coordinated with the PM.
- Delivery destination
- The agreed channel for files, status and clarification.
Handoffs and corrections
A request changes state.
Its direction stays connected.
Each transition has a clear purpose. Missing information returns to the conversation; quality corrections return to the production task that needs them.
Received
Read the request and its production requirements.
Routed
Classify the task and pass it to the right production agent.
Produced
Create the assets and versions following the supplied direction.
Checked
Review the output and route corrections when needed.
Delivered
Share checked files and a completion message.
ClarificationRequest information → Designer’s response → Continue the task
CorrectionQA finding → Responsible production agent → QA review
Technology and client context
A defined role.
A task-specific setup.
Instructions & context
The client’s design direction and production requirements inform how its agents are trained and configured.
Models & tools
The combination depends on the work: interpreting requests, operating production tools, generating assets or reviewing output.
Security & evolution
Brother Hills manages the system’s security and development as client needs and available technology evolve.
We work with technologies from OpenAI, Anthropic and ElevenLabs, open-source technologies and our own technology trained and developed by Brother Hills.
Applied AI / Research notes
Inside the
production system.
The architecture defines each role. These notes explore the AI concepts behind those roles, and the decisions that turn a designer’s request into a repeatable production process.
The context window
What does the agent need to know at this step?
A focused context.
A continuous brief.
A context window is the finite working space available to a model during an interaction. Instructions, references, conversation history and tool results occupy that space; the model also needs room to generate its response. Capacity is measured in tokens, the units used to represent input and output, and the limit depends on the model.
A larger window can hold more information. It does not, by itself, decide which artwork is current or which copy has been superseded.
Applied to production
An Adaptation task needs the approved key, current content and target formats. QA needs those same requirements alongside the produced files. Each step benefits from relevant, current context.
A conceptual view of one production step
Selected information for the current step, with capacity for the model’s response.
- Context
- Information available to the model for the current interaction.
- Persistent state
- Information retained between interactions. It must be supplied or retrieved when it is needed.
Technical referencesAnthropic / Context engineeringOpenAI / Conversation state
Interpretation & routing
What does the request mean for production?
Understanding comes
before routing.
The model supporting Intake needs to interpret a designer’s request, identify the work and recognize missing information. A useful classification separates supplied requirements from questions that still need an answer.
Applied to Intake
A single message may involve several production roles. Intake should identify the tasks and their dependencies, then pass the right information to each responsible agent.
Illustrative request → classification
“Create square and vertical versions of the approved key, using copy B, for tomorrow.”
- Work
- Adaptation + Content
- Reference
- Approved key artwork
- Content
- Copy B
- To clarify
- Exact dimensions and delivery time
If the project context already supplies a requirement, reuse it. Ask the designer only for what is still missing or conflicting.
Choosing the model for the role
Representative briefs make the choice measurable: incomplete requests, mixed tasks and conflicting versions. Interpretation, routing quality, response time and cost all matter. The model is one component; instructions, tools and coordination shape the complete agent.
Technical referencesAnthropic / Building effective agentsOpenAI / Evaluation practices
Prompting
How does direction become a usable instruction?
Clear instructions.
Visible boundaries.
A production prompt connects a defined task, the relevant references and an expected result. Separating these elements helps distinguish instructions from source material, examples and missing information.
Applied to the designer’s direction
“Adapt this campaign” leaves essential decisions unstated. Useful instructions identify what changes, what remains consistent, which material is authoritative and when clarification is needed.
- Instructions
- Describe the task, constraints, expected output and handling of uncertainty.
- Source material
- Contains the artwork, copy and references to work with. Text inside an asset is not automatically an instruction.
Read an illustrative Intake instruction
- Identify the work.Classify the production tasks in the designer’s message. Use the current project brief to resolve references.
- Keep the supplied requirements.List the artwork reference, content version, formats and deadline. Preserve explicit design direction.
- Make uncertainty visible.Flag missing or conflicting requirements. Do not invent dimensions, copy or delivery times.
- Return a usable handoff.Provide the task classification, source references, output requirements and any clarification needed before routing.
Prompting, configuration and training
Changing a prompt changes the instructions the model receives. Configuring an agent also involves its context, tools and workflow. Fine-tuning is a separate process that changes a model using training examples. Each addresses a different need.
Technical referencesOpenAI / Prompt engineeringOpenAI / Model optimization
Agent development
How does interpretation become an action?
A defined task.
The tools to carry it out.
An agent system connects a model with instructions, context and tools. The model can request an operation; the surrounding software executes it, returns a result and coordinates what happens next.
Applied to Adaptation
The task starts with approved artwork and target formats. Development must account for opening the correct source, producing the requested versions, exporting the assets and handling a missing file or failed operation.
A development cycle for a production role
Specify inputs, output requirements and what successful completion means.
Provide the tools and scoped access required to carry out that task.
Test complete briefs, missing assets, conflicting instructions and operation failures.
Compare the revised workflow against the same cases before expanding its scope.
Coordination matters
A tool response is evidence of an operation, not proof that every requirement was met. Completion needs to connect the produced files with the original task, QA findings and the intended delivery destination.
Technical referencesAnthropic / Tool useAnthropic / Building effective agents
Evaluation & evolution
Does the system remain reliable as it changes?
Test the decisions
as well as the files.
QA checks a production output against its brief. Evaluations examine how the system behaves across a set of cases, including after changes to prompts, models or tools.
IntakeDoes it identify the right tasks and recognize missing requirements?
ProductionDoes it preserve the supplied design direction, content and requested formats?
QA & correctionsDoes it identify a mismatch and return the correction to the responsible agent?
DeliveryDoes it communicate the checked files and completion status to the correct project channel?
Applied to the complete workflow
A correct-looking file can conceal a missed requirement or an incorrect handoff. Useful evaluation follows the full request and compares requirement coverage, correction outcomes, elapsed time and cost. These are evaluation criteria, rather than published performance results.
Technical referencesOpenAI / Evaluation practicesAnthropic / Agent evaluations
Let’s talk
Let’s look at the work
behind your workflow.
Tell us about your production requirements and the systems your team works with.