Page header

COSMIC AI Agent Framework

Six design lenses for production-grade AI agent systems

Building an AI agent that works is becoming increasingly easy.

Building an agent system that continues to work reliably, securely, and economically when it becomes part of real business processes is a different challenge.

A prototype can perform impressively with a good prompt, access to several tools, and the right model. But as agents start interacting with business systems, making decisions, triggering workflows, sharing context, handling sensitive information, and collaborating with people or other agents, the system around the agent becomes just as important as the intelligence inside it.

COSMIC is Chapman Bright’s framework for thinking about that transition.

It provides six design lenses for moving from isolated agent experiments toward production-grade agent systems:

Composability · Optimize · Secure · Monitor · Improve · Cost

COSMIC is not a prescribed technical architecture. Different agents require different levels of autonomy, control, sophistication, and governance. Instead, the framework helps teams ask the questions that matter when agents move from: “It works” to “It scales.”

C | Composability

How should the system be structured?
The first design question is not how much an agent can do.
It is how the wider system should work together.

A common starting point is one agent with an increasingly large set of instructions, tools, and responsibilities. This can work for a prototype, but complexity grows quickly as more capabilities are added.

Production systems often benefit from separating responsibilities across:

  • Specialized agents
  • Reusable sub-scenarios and workflows
  • Deterministic automation
  • Tools and APIs
  • Shared data and state
  • Human decisions and approvals

The objective is not to turn everything into an agent. It is to determine what should reason, what should execute, and how those components should work together.

For critical processes, this can mean allowing an agent to decide what needs to happen while a controlled deterministic workflow determines exactly how it happens.

This can improve reliability, reuse, governance, and maintainability without removing the flexibility that agents provide.

Questions to consider:

  • Should this responsibility belong to the main agent, a specialized agent, or a workflow?
  • Which capabilities need to be reusable?
  • How do components exchange context, IDs, and state?
  • Which actions should be directly available as tools?
  • Which actions should sit behind a controlled deterministic process?

O | Optimize

How do we make it perform better?
A more powerful model does not automatically create a better agent.

Performance depends on how models, instructions, context, knowledge, tools, and architecture work together.

An agent with excessive context may become slower and less focused. A very capable model may add latency where a smaller model would perform the task equally well. Poor tool descriptions can make even a strong model choose the wrong capability.

Optimization is therefore about giving each part of the system what it actually needs.

Important areas include:

  • Context versus knowledge versus instructions
  • Knowledge-file design
  • Model selection based on quality and speed
  • Conversation-history management
  • Clear tool descriptions
  • Clear input and field descriptions
  • Parallel versus sequential execution

The objective is not maximum intelligence. It is the right intelligence, information, and execution pattern for the work being performed.

S | Secure

How do we keep it within safe boundaries?
As an agent gains access and autonomy, the consequences of a wrong decision can increase. There is a meaningful difference between an agent that recommends an action and an agent that can execute it. Security therefore needs to be designed into the system.

Relevant considerations include:

  • Prompt and input guardrails
  • Human-in-the-loop approval
  • Least-privilege access
  • Limiting the blast radius of failures
  • User identity and permissions
  • Sensitive-data handling
  • Regulatory and disclosure requirements
  • Controlled access to critical tools and workflows

The goal should not automatically be maximum autonomy. The goal is the appropriate level of autonomy with the controls required for the business process. Human judgment remains important where quality, accountability, relationships, or irreversible actions matter.

M | Monitor

How do we know what it is doing?
A successful output shows that an agent worked once. Production requires knowing whether the system continues to work. As agent systems become more complex, teams need visibility into what happens across the complete chain.

That can include:

  • Agent and workflow executions
  • Tool calls
  • Errors and failures
  • Latency
  • Token usage
  • Model usage
  • Decision paths
  • Evaluation results
  • Unexpected behaviour

Observability allows teams to move from: “It seems to work.”
to: “We understand how it is working.”

Without that visibility, failures can remain hidden and improvements become difficult to prioritize.

I | Improve

How do we make it better over time?
Production is not the end of development. Models change. Knowledge changes. Processes change. Tools evolve. Real usage exposes situations that were not anticipated during initial testing. Agent systems therefore need a structured improvement loop.

This can include:

  • Defined test cases
  • Regression testing
  • Output evaluation
  • LLM-as-judge patterns
  • Versioning of system instructions
  • Versioning of agent and scenario configurations
  • Change documentation
  • Controlled releases
  • Rollback capabilities

Over time, parts of the improvement process can themselves become more intelligent.

For example, evaluation results, failures, and previous agent versions could be made available to an improvement agent that identifies potential changes and prepares them for review.

The principle remains: Observe → evaluate → improve → test → version → repeat.

Improvement should be controlled and evidence-led rather than a sequence of untracked prompt changes.

C | Cost

How do we keep it economically sustainable?
An agent system can work technically and still fail economically. Every model call, token, file, retrieval action, tool execution, and repeated workflow contributes to operating cost. That makes cost an architectural concern, not just a procurement concern.

Relevant considerations include:

  • Model selection based on cost
  • Context-window usage
  • Conversation-history length
  • Using text instead of files where appropriate
  • Avoiding unnecessary LLM calls
  • Using deterministic execution where reasoning adds no value
  • Attributing AI usage to specific agents or capabilities
  • Separating provider projects where useful for cost tracking
  • Monitoring cost relative to business value

A production system should eventually answer not only: What does this agent cost?
but: What value does this agent create relative to that cost?

Scalability includes economic sustainability.

From isolated agents to systems of work

COSMIC reflects a broader shift in how we think about AI agents. Many early use cases treat an agent as a standalone virtual employee: one agent, one responsibility, one collection of tools. But business processes rarely work in isolation.

Real work crosses:

  • People
  • Teams
  • Systems
  • Data
  • Workflows
  • Decisions
  • Controls

Agents need to become part of that environment. That means designing beyond the individual agent. COSMIC helps teams look at the wider system through six connected lenses:

Lens Core question
Composability How should the system be structured?
Optimize How do we make it perform better?
Secure How do we keep it within safe boundaries?
Monitor How do we know what it is doing?
Improve How do we make it better over time?
Cost How do we keep it economically sustainable?

 

Not every system requires the same answer. A bounded internal assistant and an autonomous agent operating across financial systems should not have the same architecture or controls. COSMIC provides the lenses. The design still needs to fit the work.

The objective is not more agents

At Chapman Bright, we do not start with the question:
Where can we use AI agents?

We start with:

How should this work, and what combination of people, processes, data, automation, AI, and agents will improve it?

Sometimes an agent is the right answer. Sometimes deterministic automation is better. Often the strongest solution combines both. The goal is not to build increasingly sophisticated agents. The goal is to design better systems of work around them.

Discover how we can help you automate.