AI agents fail for the same reason untrained falcons fail: they follow instinct, not intent. After deploying 150+ voice agents handling millions of banking calls at Cake, I've found the fix isn't better models—it's better context engineering. Here's the framework.

In UAE, master falconers earn millions. They spend months building trust with a bird that could fly away forever. Two GPS trackers on every falcon—not because the bird is malicious, but because it follows instinct, not intent. You have a tool of terrifying capability. 300km/h dives. Genius-level pattern matching. Also: easily fooled by decoys, attacks wrong targets, zero clue why you want what you want.

Why AI Agents Attack Wrong Targets

The most common agent failure isn't capability—it's context. A generalist agent seeing everything attempts everything. At Cake, our early voice agents had a 34% misrouting rate. They'd hear "transfer" and initiate wire transfers when customers meant "transfer me to a human."

Falconers figured this out centuries ago: you don't say "go hunt." You choose terrain. Read wind. Pick target type. Time the release.

The fix was domain isolation. We route intents to specialized agents, not one general agent. A balance-inquiry agent never sees wire-transfer tools. A complaint agent never sees transaction APIs. Narrowing context from "everything" to "exactly what's needed" dropped our misrouting rate to 8%.

Context engineering isn't prompt engineering. It's choosing what the agent sees before it acts—routing logic, tool access, memory scope. Agents with narrower context outperform generalist agents on resolution rate.

Context Engineering: Choosing the Terrain

Every agent deployment requires three context decisions:

1. Intent routing. Which agent handles which query? We use a lightweight classifier (not the main LLM) to route before the expensive model runs. Misclassification here cascades into everything else.

2. Tool scoping. What can this agent touch? Our payment-dispute agent can read transactions but not modify them. It escalates to a human for reversals. Fewer tools = fewer attack surfaces = fewer hallucinated actions.

3. Memory boundaries. How much conversation history? We cap at 4 turns for transactional queries, unlimited for complex disputes. Long memory invites confusion. Short memory loses context. Match memory to task complexity.

Verification Before Reward: The RLVR Pattern

Falconers never reward mid-flight. The falcon returns. You check: Right prey? Actually caught? Not someone's pet? Only then: reward.

This is RLVR—Reinforcement Learning with Verifiable Rewards. Math has right answers. Code runs or crashes. Banking transactions either reconcile or they don't.

At Cake, every voice agent call gets scored after completion: Did the customer's stated intent get resolved? Did any compliance flags trigger? Did the customer call back within 24 hours? We weight resolution (40%), compliance (35%), and repeat-call rate (25%). Agents that score below threshold get their prompts reviewed, not retrained blindly.

Never optimize agent behavior on intermediate outputs. Verify the final outcome. A confident-sounding wrong answer is worse than an uncertain right one.

Monitoring: Two GPS Trackers

Every falcon wears two GPS trackers. Redundancy. Our agent monitoring follows the same principle:

Runtime logging: Every call captures intent classification confidence, tool calls made, tokens consumed, and latency. This catches drift before customers complain.

Outcome tracking: Resolution rate, escalation rate, customer satisfaction, repeat contacts. This catches whether the agent actually helped.

The gap between runtime metrics (agent did things) and outcome metrics (things worked) is where failures hide. An agent can execute perfectly and still fail the customer.

When Humans Take Over

Our human override rate sits at 23%. That's not a failure metric—it's a design choice. Agents handle volume. Humans handle edge cases, emotional situations, and anything involving regulatory judgment.

The falcon doesn't hunt for you. It hunts with you. You bring judgment it'll never have. It brings speed you'll never have. Current AI agents are falcons, not dogs. Dogs understand intent. Falcons follow training.

We explicitly train agents to recognize their limits: "I'm not confident I can help with this. Let me connect you with a specialist." A graceful handoff beats a botched resolution.

The Framework Summary

Deploying AI agents at scale requires four things:

1. Context engineering: Route to specialized agents. Scope tools narrowly. Bound memory appropriately.

2. Verified rewards: Score on outcomes, not outputs. Resolution rate > confidence score.

3. Dual monitoring: Track both what agents do and whether it worked.

4. Designed handoffs: Know when humans take over. Build the escalation path first.

The best falconers aren't those with fastest birds. They're those who know when to release, when to recall—and honest enough to check if the catch was actually what they asked for.

Pham Thanh Lam is Chief Data Science & AI Officer at Cake by VPBank, Vietnam's largest digital bank, where he leads a team running 100+ AI models and 150+ voice agents in production.