Guardrails
In Scrum Manager terminology, guardrails are the technical implementation of the Always / Ask First / Never system: the Always constraints (what the agent can always do), the Ask First constraints (what requires confirmation), and the Never constraints (what is prohibited) translate into guardrails in the agent's configuration.
Types of guardrails
Content guardrails
Restrictions on what types of outputs the agent can generate: do not produce code that accesses secrets, do not include personal data in logs, and do not generate content outside the task's domain.
Action guardrails
Restrictions on what actions the agent can carry out autonomously: not committing to production branches, not sending external communications without approval, and not modifying environment configuration files.
Scope guardrails
Restrictions on which parts of the system the agent can touch: access limited to certain directories, read-only access to production databases, and a ban on modifying dependencies without approval.
Spending guardrails
Restrictions on resource consumption: a token limit per session, a limit on calls to external APIs, and an alert when costs exceed a threshold.
Guardrails and AI governance
Guardrails are the technical layer of AI governance. Governance defines the policies (what AI can and cannot do); guardrails implement them in the system. Without guardrails, governance policies are documents with no technical effect. Without governance, guardrails are arbitrary restrictions with no grounding in the team's values and goals.
Guardrails in the agile context
In the SDD framework, guardrails are specified in the Boundaries section of the spec: the constraints that apply to the agent while implementing that particular spec. Unlike global guardrails (which apply to all of the team's agents), spec guardrails are local to a specific task.
Common mistake
Defining guardrails only for the most obvious actions and forgetting about side effects. An agent restricted from "modifying production files" can still compromise security if it is able to read production secrets and write them to a log file. Effective guardrails are designed in terms of risk, not individual actions: what is the worst damage the agent could cause with the permissions it has?
See also
Want to advance in agility? You can search for course and exam calls or go at your own pace by becoming a member of the Agile Club. This membership includes exclusive resources, e-learning classrooms, and access to Skill Arena: a space to practice and measure your agile skills at your own pace.