Agentic Work Patterns
Why I'm Building This
AI agents can help in very different ways. Some behave like configured assistants or follow a bounded workflow. Others can retrieve information, use tools, take approved actions, or work toward a goal with more autonomy. What matters to me isn't how autonomous they sound. It's whether they improve real work without making responsibility harder to see.
Over my career, I've seen important workflows depend on people repeatedly reconstructing the same expectations: what information matters, who owns a recommendation, when a different communication channel would be safer, and what someone needs to understand before acting.
I created Agentic Work Patterns to capture lessons like these in a form that other people can understand, test, and adapt. Some patterns will draw on earlier work, but they won't copy private implementations. I'll rebuild them from scratch, remove organization-specific information, and test them with fictional material.
Other patterns will come from what I'm learning now as I explore building an AI-native startup. They may support analysis, communication, coordination, creation, operations, decision preparation, or something else entirely. Decision-Ready Escalations is the first example, not the boundary of the collection.
The GitHub repository is the public home for these patterns as they become ready to share.
What I Mean by an Agentic Work Pattern
To me, an agentic work pattern starts with a recurring work problem—not with a model feature or a particular platform. It doesn't have to involve a fully autonomous agent. The right interaction model, tools, actions, and level of autonomy depend on the work being improved.
Each pattern should answer five basic questions:
- What problem is it trying to solve?
- What should the agent do?
- What information, tools, and actions can it use?
- When should it stop or hand the work back to a person?
- What evidence suggests that it helps, and what remains unproven?
The pattern should also remain separate from the product used to implement it. A Microsoft 365 Copilot version may show one way to build the pattern, but the product isn't the pattern.
Current Patterns
| Pattern | What it helps with | Human boundary | Explore |
|---|---|---|---|
| Decision-Ready Escalations | Turns incomplete or overloaded escalation notes into a brief a decision-maker can act on. | The agent prepares the handoff. The sender owns the facts and recommendation; the recipient makes the decision. | Pattern · Instructions · Evaluations · Copilot guide |
This first pattern draws on firsthand experience, but the public version was rebuilt from scratch and tested with six fictional scenarios. It hasn't been independently validated, and I'm not claiming measured productivity gains or production reliability.
The First Pattern: Decision-Ready Escalations
Decision-Ready Escalations helps someone turn incomplete or overloaded notes into a clear brief. The agent looks for missing context, organizes the options and trade-offs, and prepares the handoff to the person who needs to decide.
It's deliberately bounded rather than fully autonomous. It doesn't validate the facts, send the message, make the recommendation, or take the final decision. For the story behind the pattern and the problem that led me to build it, read the article:
The GitHub pattern contains the reusable specification, evaluations, and implementation guidance.
How This Collection Will Evolve
I'll add patterns over time as I revisit useful lessons from earlier work and test new approaches while building. They won't all be about decisions or escalations, and they won't all use the same level of autonomy. What they should share is a clear problem, explicit boundaries, visible human responsibility, and enough evidence for someone else to evaluate them honestly.
That evolution has already started. Since the first release, I've clarified the repository's agent terminology, strengthened its human-oversight boundaries, added optional recipient-familiarity behavior, expanded the test coverage, and made it clearer that the examples are fictional. I expect the collection to keep changing as I learn from using and testing it.
I'll keep this page short. When I add or significantly revise a pattern, I'll update the description and links here. The full specifications, implementations, and evaluation history will remain in GitHub.
Explore the Repository
Visit Agentic Work Patterns on GitHub to explore the patterns, principles, fictional test cases, evaluations, and implementation guides.