[0:00] Today, there's a lot of excitement around AI agents.
[0:03] We've seen impressive demos of agents that plan, reason,
[0:13] and act across tools.
[0:16] But the real question isn't whether we can build them or not.
[0:20] The real question is what it takes to make an AI agent effective in real-world environments.
[0:26] When agents move from demo into production systems,
[0:30] many fall short, not because technology is incapable,
[0:34] but because real-world problems are complex, constrained, and interconnected.
[0:40] So instead of focusing on what AI agents are, I want to focus on how they behave in practice
[0:46] when embedded into real world systems.
[0:49] Most real world agent problems share the same core challenges.
[0:54] The first one, they span across multiple systems.
[1:00] Second one, they involve a lot of policies, approvals, and rules,
[1:06] they must fit into existing workflows,
[1:11] and the human should always be in the loop.
[1:16] Because of this, successful agents aren't standalone decision makers.
[1:21] They act like coordination layers,
[1:24] maintaining context, orchestrating actions across systems, enforcing
[1:29] rules, and determining when control needs to be transitioned to a human.
[1:35] A common pattern in real-world agent system is coordinating a sequence of
[1:40] actions across multiple systems while managing state,
[1:48] timing,
[1:51] rules, and exception.
[1:56] This pattern shows up anywhere a single event triggers a multi-step workflow with dependencies.
[2:04] One concrete example of this pattern is onboarding a new employee.
[2:11] Onboarding isn't an easy task.
[2:14] It's a workflow composed of many steps, starting with provisioning, access and entitlements,
[2:21] ordering required resources, scheduling initial activities, assigning required trainings and tracking them to completion.
[2:32] In this use case, agents don't replace people.
[2:37] It uses context-based signals such as roles, location, and start date to sequence actions across systems,
[2:47] monitor workflow state, and flag deviation from expected behavior.
[2:52] The hard part isn't reasoning.
[2:54] It's reliably orchestrating multiple systems while respecting policy and timing constraints.
[3:02] Another recurring pattern is policy-governed action execution, where risk, rules, and access control shape
[3:11] what actions a system is allowed to take.
[3:14] This pattern appears whenever a system is handling incoming requests with very level of sensitivity or impact.
[3:23] IT support is a good representation of this pattern.
[3:26] In this case, agent may process requests such as passwords, software
[3:33] or hardware resources, any requests that come through,
[3:40] ticketing, and routing of any requests.
[3:45] Some requests follow a well-defined and low-risk execution path.
[3:50] Others require validation, approval, and sometimes escalations.
[3:55] An effective agent in this case interrupts requests intent, evaluates the applicable policies,
[4:04] automatically executes some of the permitted actions, escalates any ambiguous or high-risk cases.
[4:14] This shows the explicit control boundaries.
[4:18] The system behaves predictably and humans step in precisely where the rules need them to.
[4:25] In other cases, agents operate inside a well-defined processes where exceptions are the real challenge.
[4:33] This pattern shows up in systems such as invoice processing or order management.
[4:39] In this case, an agent may,
[4:41] extract
[4:45] structural data, match it against the existing record, validate it
[4:56] against rules or concerns,
[4:58] or route approval and lastly, update the downstream systems.
[5:08] This is a happy path.
[5:11] Which is straightforward.
[5:13] The real complexity lies in handling missing data, mismatch data, or any non-standard conditions.
[5:24] Agents add value by consistently handling predictable flows and surfacing only through exception for human reviews.
[5:33] Another important pattern involves triaging and routing large volumes of incoming work.
[5:40] This pattern appears wherever the system needs to prioritize attention under load.
[5:46] A customer service is a great example for this.
[5:50] Here an agent
[5:52] must analyze
[5:58] and categorize incoming requests.
[6:02] Route
[6:05] work to the appropriate teams,
[6:07] and suggest responses based on historical data.
[6:15] Humans still resolve the issues, but agents ensure the priority, context, and routing decisions are applied consistently at scale.
[6:26] The pattern holds regardless of where the work originates.
[6:30] Across all the patterns that we saw, regardless of the domain, the same characteristics apply.
[6:38] A successful AI agents are narrowly scoped, they orchestrate across systems, apply rules, and relate its signals.
[6:56] Keep human in the loop and are designed for integration and not isolation.
[7:05] These systems don't feel like flashy AI features.
[7:09] They feel like well-designed components of a larger architecture.
[7:13] The real power of AI agents is in the autonomy.
[7:17] It's its alignment with real workflows, limits, and control structures.
[7:23] When agents are designed around coordination, rules and accountability.
[7:29] They stop being experiments and start operating as reliable components in production systems.
[7:36] That's what it takes to make AI agents work in the real world.