Cedric Clyburn outlines five concepts for using coding agents, models that run multi-step software tasks with tools. His ten-minute IBM Technology lesson moves from requirements to architecture, model limits, security, and developers’ future work.

Specify the outcome (00:53)

  • Clyburn describes his team’s shift from autocomplete to assigning entire features, including codebase exploration, tests, and changes.
  • Spec-driven development records requirements and expected outcomes before the agent begins. It gives the work a clearer target than repeated prompt negotiation.
  • Define what completion means, as well as what the software should do.

Conduct the work and design the loop (02:28)

  • Developers still choose architecture and divide goals into milestones. Those decisions shape what agents implement.
  • Context engineering manages the model’s limited working input. Code, tool descriptions, reusable instructions, and external data all compete for space.
  • Loop engineering builds the system that repeatedly advances the task, instead of relying on a human to send every follow-up prompt.

Learn the model’s operating limits (04:52)

  • A model’s estimate of project duration can differ sharply from what a coding agent produces. Clyburn illustrates this with a game prototype.
  • Model weights occupy GPU memory. The KV cache, stored intermediate attention data, also uses memory as the model processes requests.
  • Skills provide reusable instructions. Scheduled or event-triggered automations start work; subagents can investigate a library and return findings to the main agent.

Contain prompt-injection damage (07:20)

  • An agent can encounter malicious instructions while reading a pull request or other external material.
  • Access to sensitive data, untrusted web content, and a way to send data creates a path for stolen credentials.
  • Clyburn recommends disposable sandboxes to reduce the damage a compromised coding agent can cause.

Plan for operating the systems (08:25)

  • Drawing on Andrew Ng’s electricity analogy, Clyburn expects continuing work on infrastructure, access, safety, reliability, and quality.
  • He argues that architectural judgment and critical thinking will matter as developers spend more time orchestrating agents.

Caveat: His productivity examples and employment outlook are observations and predictions, not controlled measurements.