iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
The four phases of AI agent maturity—Crawl, Walk, Run, and Fly—describe a shift from fixed automation to systems that can pursue goals with increasing independence. The framework is a practical model proposed by technology architect Denis Prilepskiy, not a standardized scale: an organization can benefit from a lower-autonomy phase without aiming for full autonomy.
What do the four phases measure?
The key difference is how much initiative the system has. Crawl systems execute predetermined logic. Walk systems respond when a person asks. Run systems can plan and carry out multiple steps toward a bounded goal. Fly describes agents handling a broader process with minimal human involvement.
To place a system in the framework, consider five questions: Does it follow fixed rules or plan dynamically? Must a person initiate every interaction? Can it use tools and execute multiple steps? How much work can it complete? What human oversight and governance does that scope require?
What are the four phases of AI agent maturity?
| Phase | Typical behavior | Initiative and oversight |
|---|---|---|
| Crawl | Executes rules, workflows, or model predictions for well-defined tasks. | Low initiative; people define the process and handle exceptions. |
| Walk | Generates or interprets language in response to a user’s request. | User-initiated and generally limited to the current interaction. |
| Run | Plans and performs multiple steps toward a bounded goal, potentially using tools. | More initiative; scope limits and oversight are important. |
| Fly | Coordinates work across a process with minimal human involvement. | Highest autonomy and broadest governance demands. |
Crawl: assisted intelligence
Crawl covers traditional automation and predictive systems: rule-based workflows, robotic process automation (RPA), simple chatbots, and classical machine-learning predictions. They handle repetitive, well-defined work by following fixed rules or returning model outputs. They do not dynamically plan a sequence of actions or take initiative to pursue a goal.
#1 Best Overall
Walk: generative AI assistants
Walk systems can summarize information, draft content, and answer questions in natural language. They typically respond to a user’s prompt rather than independently deciding what work to start or what steps to take next. Prilepskiy’s December 2025 article cited Microsoft Copilot in Office apps, Google Duet AI for Workspace, and custom GPT-based chatbots as examples; those are dated illustrations, not a current comparison of product names or capabilities.
Run: goal-driven AI agents
A Run-phase agent receives a high-level but bounded objective, then plans and executes a sequence of actions. It may call APIs or other tools, use memory, and adjust based on feedback. For example, an IT support agent could read a ticket, inspect logs or a knowledge base to diagnose the issue, apply a fix, check whether it worked, and escalate an unfamiliar case. The defining change from Walk is not simply more fluent conversation: it is the ability to act across multiple steps toward a goal.
Fly: autonomous agentic systems
Fly describes one or more agents handling an end-to-end process with minimal human involvement. Prilepskiy illustrates this with order fulfillment spanning inventory checks, shipping, and customer updates. In his December 16, 2025 article, he characterized this level as largely experimental or conceptual and said few organizations had anything close to it in production. That is the author’s assessment at publication, not a measured estimate of adoption or a verified account of prevalence in 2026.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteHow should an organization use the framework?
Use the phases to describe a system’s operating model and the controls it needs, not as a maturity score that every organization must maximize. A reliable fixed workflow may be a better fit than an agent where the work is predictable, the consequences of error are high, or exceptions are hard to contain.
Rank #3
- Start with the work: Identify a repetitive, valuable task and determine whether rules or a user-prompted assistant can address it before introducing multi-step autonomy.
- Bound agent goals: For a Run-phase system, define what the agent may do, which tools and data it may access, and when it must stop or escalate.
- Keep consequential decisions reviewable: Require human approval for important actions where errors could cause material harm, and retain a route for escalation.
- Scale controls with autonomy: More initiative and wider process scope increase the need for process readiness, sound architecture, governance, and risk controls.
Prilepskiy recommends moving incrementally—“walk before you run (and certainly before you fly).” He argues that a Run-phase agent with a human overseer can offer a useful balance between efficiency and risk management; this is practical guidance from the framework’s author, not a tested outcome that applies to every organization.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the framework does—and does not—tell you
The phase labels help distinguish rule execution, prompted assistance, multi-step goal pursuit, and process-level autonomy. They do not establish that an organization should reach Fly, define a formal certification threshold, or provide a validated way to rank companies. The original article’s broad statements about where companies cluster and how rare Fly deployments are were not accompanied by a named survey or figures, so they should not be treated as current industry statistics.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
