Long AI coding sessions often degrade as the context window fills. The agent starts strong, then begins losing track of earlier decisions, drifting from constraints, producing circular fixes, or expanding the diff in ways that no longer match the original task. This degradation is a workflow risk that can create unreliable output and hidden rework.

This talk introduces context rot as a practical engineering problem in agentic development. Teams need disciplined ways to break work apart, manage session boundaries, and preserve continuity without polluting future sessions with bad assumptions. Large tasks should be decomposed into bounded steps. Each pass should have 1 job. Fresh-session handoff prompts can be more reliable than 1 long session trying to do everything.

The session explains how to structure multi-step work using orchestrator and subagent sessions. The orchestrator manages task state, coordinates handoffs, and incorporates human feedback. Subagent sessions each handle 1 bounded task with a clean context window. This pattern reduces context pollution, improves reliability, and can reduce token cost per task.

Attendees will learn how to identify signs of agent thrashing, stop and rescope safely, write handoff prompts, and capture useful understanding in durable artifacts. This is a tactical session for teams using agentic coding tools heavily and seeing quality degrade in longer sessions.

This talk directly references the author’s Engineering Standards for Agentic Software Development, especially TB1, TB3, TB4, TB5, TB6, CU7, and CU8: https://www.linkedin.com/pulse/engineering-standards-agentic-software-development-edensoft-park-ki1se

Target audience:
Senior engineers, staff engineers, AI tooling teams, platform teams, and developers using agentic coding tools for substantial implementation work.

Preferred Session Duration:
50 mins including Q&A