Explore agentic AI harness patterns and source-backed learning cards, alongside generated YouTube articles when available.
Ten reference-backed patterns from the shared visual. These are concise learning cards, not transcript-generated articles. Some patterns overlap and can be combined within one system.
An agent alternates between deciding what to do, taking an action (often a tool call), and using the observation to choose its next step.
Flow: Decide → act → observe → repeat or finish
Use when: A task needs several tool calls and the next action depends on real-world results.
Example: A research agent searches, reads a result, and decides whether another search is needed.
Classifies an incoming request and sends it down an appropriate specialist path, prompt, tool, or model.
Flow: Request → classify → selected path
Use when: Requests fall into distinct categories that benefit from different handling.
Example: A support assistant sends billing questions to a billing flow and bugs to technical support.
A planning step proposes subtasks; an executor carries them out and can return results for replanning.
Flow: Goal → plan → execute steps → review/replan
Use when: A multi-step job benefits from a visible plan that may change as new information arrives.
Example: A coding assistant plans a refactor, edits files, runs tests, and revises the plan after failures.
Runs independent subtasks at the same time, then combines their outputs; tasks are known before dispatch.
Flow: Task → parallel branches → aggregate
Use when: Work can be divided into independent pieces or compared across multiple perspectives.
Example: Separate reviewers check correctness, security, and accessibility concurrently.
A central agent decides which subtasks are needed, delegates them to workers, and synthesizes their results.
Flow: Orchestrator → dynamic workers → synthesis
Use when: The number and type of subtasks cannot be fully predicted in advance.
Example: A code-change agent delegates work across whichever files a request actually affects.
One model drafts an output; another evaluates it against criteria and sends actionable feedback for revision.
Flow: Generate → evaluate → revise → stop when acceptable
Use when: Quality criteria are clear and iterative critique can measurably improve an answer.
Example: A writing agent revises a translation after a reviewer flags lost meaning.
A supervisor retains responsibility for the overall task while calling specialized agents for bounded pieces of work.
Flow: Supervisor ↔ specialists → supervisor responds
Use when: One agent should own the final answer but needs domain-specific help.
Example: A customer-facing supervisor consults booking and refund specialists before replying.
An agent transfers control to another agent, which takes over the relevant part of the interaction.
Flow: Triage agent → specialist takes control
Use when: A specialist should handle the user directly rather than just return a result to a manager.
Example: A triage assistant hands a refund request to a refund agent.
Execution pauses before a sensitive action until a person approves, rejects, or changes the proposed action.
Flow: Proposed action → pause → human decision → resume
Use when: An action has financial, security, privacy, or other significant consequences.
Example: An agent requests approval before sending a refund or executing a destructive operation.
Saves workflow state at checkpoints so an interrupted run can resume instead of restarting from the beginning.
Flow: Run → persist state → interrupt/fail → restore → resume
Use when: Tasks are long-running, may wait for approval, or must recover from process failures.
Example: A customer-support workflow pauses for approval, then resumes using its saved thread state.
Aishwarya Srinivasan · Sep 2, 2026
Aishwarya Srinivasan · Sep 2, 2026
Aishwarya Srinivasan · Sep 2, 2026