Generating code is no longer the hard part. The hard part is coordinating agents so their work is verifiable, governed, and repeatable across a real engineering organization. In practice, the model is rarely the bottleneck — the harness around it is.

This session looks at agentic development from the developer’s seat and breaks down what a production-grade harness needs: planning and task decomposition, subagents with isolated context, reusable skills, tool access through MCP, and human checkpoints where judgment matters most.

Then we put it to work. In a live demo with IBM Bob, we’ll take a multi-step change across a real codebase: Bob plans the work, spawns parallel agents, and validates their output before anything is merged. From there, we’ll show how the same harness carries across IDE, CLI, and CI/CD pipelines, and how teams track cost and impact as usage grows.

You’ll leave with:

A practical model for what belongs inside an agent harness

Patterns for delegating work to agents safely

A framework for evaluating harness choices in your own organization