The AI Skills That Turn a Developer into a Conductor

1. IT has never stayed the same for long — change is natural here. Assembly gave way to compilers, on-premise gave way to cloud, and each shift changed what the job required, just like AI does it now.
2. The role is shifting from writing code to orchestrating agents. Hundreds of engineers at RBI already work this way. The manual line-by-line work is being absorbed; the judgment about what to build and how remains human.
3. An agent has a goal, tools, and a loop that runs until the goal is reached. Understanding how that loop works — and where it fails — is the practical skill.
4. Start with an agent, not a course. The fastest way to develop agentic skills is to build something with an agent, run into its limitations, and learn from what breaks. Frustration at the start is part of the process.
5. Stay curious without chasing everything. The field moves fast enough that the next big thing is sometimes irrelevant three weeks later. Missing two weeks does not matter. A consistent habit of reading, trying, and experimenting does.

The anxiety about AI changing or replacing software development jobs is real, but Philip places it in a longer sequence. Assembly programmers in the 1950s feared that compilers would make them redundant. Ops engineers feared the cloud would remove the need for their expertise. Neither outcome materialised fully.

"Whenever building gets easier, more and more stuff will be built," he says. The layer of abstraction that AI provides does not compress the field — it expands what is possible to create and shifts where the human contribution sits.

Philip uses a conductor metaphor: in an orchestra, the conductor does not play an instrument. They hold the full picture of what the piece should sound like, understand what each section is capable of, and direct the ensemble toward a result that no single player could produce alone.

In software development, the equivalent shift is already happening. Agents handle the manual work — writing code, finding bugs, generating documentation — while the developer's role moves toward defining goals, evaluating outputs, and deciding when the result is good enough.

In March 2023, RBI ran a GPT hackathon. Philip had just joined as a solution architect, and his onboarding project, assigned mostly to help him explore the bank's environments, was to build something with the new technology. Chat GPT was recently introduced and proved itself useful — why not build a secure, compliant version for RBI?

He started building in March 2023. The first test users were on the service by April. Today, more than 20,000 users access RBI GPT regularly. The combination of genuine curiosity, willingness to safely experiment in a yet uncertain environment, and enough technical understanding produced something with real organizational impact – quickly.

An agent combines four things: an LLM as the reasoning core, tools it can use to take actions, a defined goal, and an iterative loop — plan, act, observe, then either finish or start again. Understanding how that loop works, where it tends to fail, and how to define goals precisely enough for an agent to pursue them usefully is what the new baseline looks like, for developers and increasingly for non-developers too.

The practical starting point is not taking a course but building something. Philip's recommendation for anyone who wants to develop these skills is to create an agent, run it on a real task, and learn from where it breaks. Early frustration is normal — the agent's capabilities are not obvious until you have pushed against their limits.

The field moves fast enough that trying to track every development is a trap. A tool that dominates the conversation in one week can be superseded three weeks later. Philip's advice is not to follow everything, but to maintain a consistent habit of paying attention to news and trying new things. The people who stay relevant in this environment are not the ones who track every release, but those who keep experimenting.