Cedric Clyburn gives an 11-minute introduction to agentic engineering on IBM Technology, from communicating with a model to operating a team of agents.

Prompts and context do different jobs

  • Prompt engineering: give the model clear instructions, constraints and examples grounded in knowledge of the task.
  • Clyburn recommends example questions and answers for smaller models; prompting choices still need testing on the actual model.
  • Context engineering: choose everything the model sees, including instructions, documents, code and available tools.
  • Long context can become less effective before the advertised limit; summaries and focused subagents help manage that working space.

The harness runs actions and returns evidence

  • A harness is the software around the model that executes tools, manages state and returns results.
  • Feed back concrete checks: did the tests pass, or did the proposed change introduce a merge conflict?
  • Sandboxes restrict what an agent can touch; orchestration coordinates its specialist workers.

Loops and graphs organize the work

  • The basic loop is goal → action → observation → decide whether to stop or continue.
  • Schedules, pull requests, bug reports and alerts can trigger work; human approval remains a separate checkpoint.
  • A graph connects several workers, such as parallel research agents handing findings to an implementation agent.
  • Extra workers consume extra tokens; use them where their contribution justifies the cost.

Software factories need ordinary operational controls

  • Specialized agents handle parts of the development lifecycle, including task management, implementation and testing.
  • Each needs isolation, an identity with bounded access, and traces showing what happened while people were away.
  • Clyburn connects these ideas to established feedback loops, graphs and distributed systems.
  • His opening claim about teams having AI write 80% or more of their code has no supporting study in the video.

“If you’re just a solo developer, you don’t need a software factory.” — Cedric Clyburn, 11:07