Operational practices for running AI agent systems in production.
Maintained by engineers working on production AI infrastructure.
Agent systems introduce behavioral change outside traditional deployment pipelines.
In traditional software systems, production behavior changes primarily through code deployment. Agent systems break that assumption.
Production behavior can change through artifacts that live outside application code:
-
prompts
-
tool routing rules
-
delegation policies
-
evaluation thresholds
-
memory configuration
These artifacts often evolve independently of application releases.
When this occurs, teams lose key operational properties:
-
controlled change
-
traceability
-
reliable rollback
This repository documents operational practices that restore those properties for agent systems.
This repository collects emerging operational practices for agent systems, including:
-
behavior admission and versioning
-
rollback and recovery for agent behavior
-
delegation control
-
tool execution guardrails
-
evaluation and regression testing
-
observability for agent decision flows
The focus is on operating agent systems reliably in production environments.
The diagram below illustrates the operational boundary where behavioral changes must be admitted before affecting production systems.
flowchart LR
A[Prompt Editing] --> B[Behavior Admission]
C[Tool Routing Changes] --> B
D[Delegation Policies] --> B
E[Evaluation Thresholds] --> B
B --> F[BehaviorSpec]
F --> G[Agent Runtime]
G --> H[Tool Execution]
G --> I[Delegation to Agents]
G --> J[Memory Access]
/practices
Operational practices for running agent systems
/specifications
Formal artifacts such as BehaviorSpec
/case-studies
Failure modes observed in agent systems
/research
Curated research and engineering references
/templates
Reusable templates for specifications and practices
Level 0 – Experimental agents
Level 1 – Controlled prompt development
Level 2 – Behavioral versioning
Level 3 – Operational guardrails
Level 4 – Controlled production systems