Agentic AI
Agentic AI architecture: the patterns, the parts and the control points
A plain guide for engineering and risk leads designing an agentic system. Pattern names come from Anthropic as read on 10 November 2026; the control points are Kriv AI practice.
An agentic AI architecture is the design that connects a language model to tools, data, memory and review steps so it can carry out multi-step work. Anthropic's article Building effective agents, read on 10 November 2026, separates workflows, where models and tools are orchestrated through predefined code paths, from agents, where models dynamically direct their own process and tool use. It names five workflow patterns: prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer. For regulated work, add control points around any pattern: permissions, human approval, logging and tests.
plain answer
What is agentic AI architecture?
It is the structure around the model: what tools it can call, what data it can read, what it remembers, who approves its actions and what is recorded. The model is one part of the system, not the whole of it.
Anthropic's article draws a line between two kinds of system. It says workflows are systems where LLMs and tools are orchestrated through predefined code paths, and agents are systems where LLMs dynamically direct their own processes and tool usage.
Source: Anthropic, Building effective agents (https://www.anthropic.com/engineering/building-effective-agents), read 2026-11-10; the page is dated 19 December 2024.
how to
Which workflow patterns does Anthropic name?
Anthropic names five. Prompt chaining decomposes a task into a sequence of steps, where each model call processes the output of the previous one. Routing classifies an input and directs it to a specialized follow-up task.
Parallelization runs subtasks at the same time, either as independent subtasks or as the same task run several times for diverse outputs. In orchestrator-workers, a central model dynamically breaks down tasks, delegates them to worker models and synthesizes their results. In evaluator-optimizer, one call generates a response while another provides evaluation and feedback in a loop.
Source: Anthropic, Building effective agents (https://www.anthropic.com/engineering/building-effective-agents), read 2026-11-10; the page is dated 19 December 2024.
how to
How should a team choose between a workflow and an agent?
Use the simplest design that does the job. A fixed workflow is easier to test, explain and audit because its path is known in advance. An agent suits work where the steps cannot be listed ahead of time, and it needs more care because its path varies.
Anthropic's article is a primary source for the pattern names. The advice on when to choose each is Kriv's practical view.
security
Which control points belong in the architecture?
Permissions: give each agent only the tools and data its task needs. Approval: require a person to approve actions that cannot be undone, such as sending, paying or changing a record. Logging: keep a record of each step, tool call and decision so it can be reviewed.
Testing: run realistic cases before launch and after any change to a model or tool. Fallback: define what happens when the agent is unsure or a tool fails. These are Kriv's controls, informed by common risk practice; they are not drawn from the Anthropic article.
kriv role
Where Kriv helps
Kriv AI designs and governs agentic systems for mid-market and regulated teams, with approval steps and records built in from the start. See our pages on agentic AI governance and on agent orchestration platforms.
Kriv AI helps teams scope, build and govern this work with human review where it matters. Kriv does not replace the vendor documentation cited on this page, and your own owners hold the final say on what goes live.
limits
What this page could not confirm
This page read one Anthropic article, dated December 2024, so some of its detail may be out of date. We did not read other vendors' architecture guides, did not test any pattern and cite no performance figures. Check the current vendor pages before you design.
Straight answers
Frequently asked questions about Agentic AI architecture: the patterns, the parts and the control points
What is agentic AI architecture?
The design that connects a model to tools, data, memory and review steps so it can do multi-step work.
What is the difference between a workflow and an agent?
Anthropic says workflows orchestrate LLMs and tools through predefined code paths, while agents let the LLM dynamically direct its own process and tool use (read 10 November 2026).
What workflow patterns does Anthropic describe?
Prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer (Anthropic, read 10 November 2026).
Where does human approval belong?
Before actions that cannot be undone, such as sending, paying or changing a record. This is Kriv practice, not an Anthropic rule.
Can Kriv design an agentic system for us?
Kriv AI designs and governs agentic systems with approval and logging built in. Your own owners hold the final say.
Talk to the team that would do the work
Bring your requirements to a working session with the person who'll actually deliver.
Book a Discovery Call