Agentic AI
Agentic AI is the ability of an artificial intelligence system to act autonomously toward a goal—making sequential decisions, using external tools, and adapting its behavior based on intermediate results. Unlike a language model that answers a single question, an AI agent can plan, execute multiple steps, and persist until it completes a complex task.
Agentic AI represents the evolution of how we use AI: from generating text on request (a single interaction) toward delegating entire tasks (multiple coordinated actions).
Characteristics
- Planning: it breaks a goal down into subtasks and determines the order of execution.
- Tool use: it can call APIs, run code, read and write files, and search the internet.
- Memory: it maintains context across steps, remembering earlier decisions and intermediate results.
- Autonomy: it operates without human intervention at every step (except at HITL/HOTL points).
- Adaptation: it adjusts its behavior when a step fails or produces unexpected results.
Types of agentic systems
- Single-task agents: designed for one specific type of task. Example: an agent that generates unit tests for any function it's given.
- Workflow agents: they orchestrate multiple subtasks with conditional logic. Example: an agent that takes a story, generates the spec, writes the code, and opens the pull request.
- Multi-agent systems: several specialized agents collaborate under an orchestrator. Each agent handles one part; the orchestrator manages the flow.
In agile teams
Adopting agentic systems transforms the workflow of Product Builders:
- Code generation, scaffolding, test creation, and API documentation are natural candidates for agentic automation.
- The developer's role shifts toward precisely defining goals (spec writing), reviewing the output, and making decisions at the control points.
- Multi-agent systems require explicit governance: what each agent can do, with which resources, at what level of autonomy, and who is accountable for its output.
Specific risks
- Error accumulation: an early error propagates and amplifies in the steps that follow.
- Chained hallucinations: an agent that hallucinates at one step makes later decisions based on incorrect information.
- Lack of context: the agent doesn't have the organizational, strategic, or relational context that the human team has.
Common mistake
Granting access to powerful tools without setting explicit limits. An agent with access to a code repository, a production database, and deployment permissions can cause significant damage if it acts incorrectly. The principle of least privilege—giving the agent only the permissions it needs for its task—is as important in agentic AI as it is in traditional systems security.
Recursos
🏦 AI applied to agile workSkill Arena · Scrum Manager
See also
Want to take your agility further? You can look for upcoming course and exam dates, or go at your own pace by becoming an Agile Club member. Membership includes exclusive resources, e-learning classrooms, and access to Skill Arena: a space to practice and measure your agile skills whenever it suits you.