Teaching AI Agents How Senior Engineers Actually Work
· CX Pulse
Most AI coding agents ship with general-purpose smarts and zero opinions about how good engineering actually gets done. They can generate code, but they don't k
Most AI coding agents ship with general-purpose smarts and zero opinions about how good engineering actually gets done. They can generate code, but they don't know your team's workflow for reviewing it, testing it, or deciding when it's ready to go live. That gap matters more than the model's raw capability.
The Pattern Worth Examining
Teams plug an AI coding agent into their stack, watch it produce output fast, and then spend weeks cleaning up after it. The agent doesn't know about quality gates. It doesn't follow the branching strategy. It skips the checks a senior engineer would run before pushing anything.
The result is speed without standards, which just creates problems faster. This same dynamic shows up in customer experience operations when you scale agents (human or AI) without encoding the actual decision-making process that experienced people use. You get volume. You lose consistency. Customers notice.
What the Agent-Skills Project Gets Right
Addy Osmani recently published agent-skills, an open-source collection of production-grade engineering skills for AI coding agents. The core idea: skills encode the workflows, quality gates, and best practices that senior engineers use when they build software. Instead of hoping the agent figures out what "good" looks like from context, you define it explicitly.
The project treats agent behavior as something you design and maintain, with structured artifacts that can be version-controlled, reviewed, and improved over time, just like the codebase itself.
Why CX Teams Should Pay Attention
The principle transfers directly to customer-facing AI. Every CX org running AI agents faces the same question: how do you encode what your best people know into a system that runs without them?
Here's what that looks like in practice:
- Quality gates before responses go out. Your senior support reps check for tone, policy compliance, and whether the answer actually solves the problem. Your AI agent should have those same checkpoints built in, not bolted on after a bad CSAT week.
- Workflow encoding, not just knowledge base access. Knowing the answer and knowing the right sequence of steps to deliver it are different things. A refund policy is information. Knowing when to offer a refund before the customer asks is a workflow decision. That distinction is where most AI agent deployments fall apart.
- Explicit best practices instead of implicit expectations. When a new human agent joins your team, they shadow someone experienced. AI agents don't get that apprenticeship unless you build it into their skill definitions. The agent-skills approach makes this a first-class concern rather than an afterthought.
The Practical Shift
Stop treating AI agent configuration as a one-time setup task. Treat it as an ongoing discipline, the same way you'd treat your engineering CI/CD pipeline or your QA process. The skills your agents use should be reviewed, tested, and updated with the same rigor you'd apply to production code.
This means someone owns the skill definitions. Someone reviews them when customer issues shift. Someone measures whether the encoded workflows actually produce better outcomes than the defaults.
If nobody owns that loop, your AI agents are running on stale assumptions fast. The agent-skills repo demonstrates this for software engineering. The same architecture applies to CX: define what good looks like, encode it as a repeatable skill, put a quality gate in front of it, and review regularly.
That work compounds. It's the difference between an AI agent that handles tickets and one that actually resolves problems the way your best people would.
What to Do With This
If you're running AI agents in a customer-facing capacity, take stock of what's actually encoded versus what's assumed. List the top ten decisions your best agents make that your AI agent currently guesses at. Those are your first skill definitions. Write them down, make them explicit, and put review dates on them.
The teams that get this right will be the ones that took the time to define what "good" means in their specific context and kept that definition current, regardless of which model they're running.
Running 20-200 agents and want to find out where your AI workflows are guessing instead of following real quality gates? CX Pulse offers a CX Friction Audit built for exactly this. DM us for pricing.