03 / AGENT SYSTEMS
Agent & Copilot Systems
We assign repeatable steps to agent systems and connect critical decisions to human approval based on permissions and risk thresholds.
Agents execute defined steps in approved tools, track state, and return control to a person at a defined risk threshold. Pricing, commitments, legal language, and customer-data decisions pass through human approval; permissions, reversibility, and auditability are part of the product.
Role-based copilots gather information, prepare drafts, and propose actions while leaving the final decision to the user. They work in the team's own context and tools rather than adding another disconnected application.
Place in the method
- 01Map
- 02Prioritize
- 03Prove
- 04Build
- 05Embed
- 06Compound
FLOW
How it runs
Each stage leaves an artefact behind; the next one builds on it.
- 01
The decision boundary
Which step belongs to the agent and which to a human: the authority matrix is written on day one with the owner of the work. Pricing, commitments, legal language, and customer data stay with people.
ARTEFACTAuthority matrix
- 02
A controlled flow
The agent works from controlled inputs: house language, permitted sources, defined tools. A guardrail is not a filter added afterwards; it is the flow itself.
ARTEFACTAgent flow
- 03
The approval experience
Human approval is designed as one step inside the work, not a separate queue. If approval becomes the bottleneck, the design is wrong and gets fixed.
ARTEFACTApproval flow
- 04
Trace and improvement
Every run is traceable: what the agent relied on, what it produced, who approved. Handover points and the escalation path ship defined.
ARTEFACTMonitoring setup
What we produce
- Agents and copilots
- Human approval systems
- Guardrails and governance rules
- Assistant experiences embedded in the workflow
When it is the right step
- Repetitive drafting and preparation consume time
- Human approval is mandatory on risky steps
- Teams need role-specific assistants
When it is not
If the need is access to the right knowledge rather than action, a source-citing knowledge system is enough.
Enterprise RAG & Knowledge SystemsIf the flow takes the same steps on every run, rules are enough; an agent adds unnecessary risk.
Workflow Automation & Integration
Outside the scope
- We do not hand pricing, commitments, legal language, or customer data to an agent; those steps stay with people by design.
- We do not build unattended flows: every agent flow ships with its handover points and escalation path.