When an AI agent fails, the first instinct is usually to blame the model. It hallucinated, it misunderstood, it wasn't smart enough. So teams upgrade to the latest release, rewrite the prompt, and try again.
Sometimes that helps. More often, the model was never the problem.
Today's models are remarkably capable. What they lack isn't intelligence, but the surrounding conditions that let intelligence do useful work. In my experience, effective agentic systems depend on four things a good prompt can't supply on its own: context, intent, boundaries and feedback.
1. Context: what the agent knows
An agent can only reason about the information it can reach. If pricing exceptions live in someone's spreadsheet, it will miss them. The same goes for customer history stored in a system it can't access, or business terms defined differently by each department.
Good context isn't about giving the agent everything. Flooding it with unrelated data makes the task harder, not easier. The goal is the right information at the right moment: tools that search and filter, clear definitions of key terms, and sources that are current and trustworthy. When an agent gives a confident but wrong answer, check what it could see before you check how it reasoned.
2. Intent: what the agent is trying to achieve
Instructions tell an agent what to do. Intent tells it why.
This distinction matters most when something unexpected happens. Two rules conflict, a request is ambiguous, or a situation isn't covered. An agent with only instructions has nothing to reason from, so it guesses. An agent that understands the goal, the acceptable trade-offs and the priorities can make a sensible call, or recognize that it shouldn't.
Most organizations have never written their intent down clearly. Humans have filled the gap with judgement and hallway conversations. Agents can't attend those conversations. If you want them to act in line with what the business actually wants, that intent has to be explicit, accessible and kept current.
3. Boundaries: what the agent is allowed to do
Capability without limits is a liability. Agents need clearly defined permissions: what they can read and change, which actions need human approval, and how much they can spend before they stop.
Critically, these boundaries need to be enforced in the system itself, not just described in the prompt. Telling an agent to be careful is not an access control. Real boundaries also include stopping conditions, so the agent knows when to pause, escalate or ask for help. And someone must own the agent, with the authority to change those boundaries as the work evolves.
4. Feedback: how the agent gets better
An agent without feedback is frozen at its launch-day quality. You need ways to know whether it's working. That means tests built from real examples and edge cases, telemetry that surfaces problems before users report them, and a loop for turning corrections into improvements.
Feedback also has to reach the people who understand the work. Engineers can tell you whether the agent ran. Only the people who do the job can tell you whether its answers were right.
The model is the engine, not the vehicle
A capable model is necessary, but it's only one component. Think of it as the engine. Context is the road map, intent is the destination, boundaries are the guardrails, and feedback is the dashboard. A better engine won't help if the rest of the vehicle is missing.
The teams getting real value from agents aren't necessarily using better models than everyone else. They've done the less glamorous work of designing the system around the model. That's where most of the improvement is waiting.
