Four ways an agent’s context goes wrong — IBM Technology
Martin Keen uses a database-driver migration to explain context engineering: choosing what information an agent receives at each step. His ten-minute IBM Technology walkthrough connects four failures to specific remedies. What fills the window (01:50) The task itself may be only a few lines. System instructions, project guidance, tool descriptions, message history and tool results occupy the rest. File contents, test output, logs and database schemas accumulate as the agent works. A larger window provides capacity; it does not guarantee that the model uses every detail reliably. Poisoning: validate conclusions before saving them (04:35) The migration agent misreads documentation and writes a note claiming the new driver automatically retries failed connections. It then deletes working retry code throughout the project. Check consequential conclusions against evidence before saving them as reusable facts. Keen’s example needs a connection test, not another confident note. Distraction: compact history, retain evidence elsewhere (05:20) A long history can steer the agent toward repeating previous actions instead of choosing what the current task needs. Replace the running log with a shorter account of the current task and progress. Keep detailed evidence in files that can be reopened when needed. Summaries lose information. Keen compares repeatedly summarizing a summary to making a photocopy of a photocopy. Confusion: expose fewer tools (06:03) Tool descriptions consume context even when the tools are unused. A large menu also makes selecting the correct action harder. Offer the tools needed for the current task and discover additional ones on demand. Keen cites a benchmark where one task failed with 46 tools and passed with 19. That example supports testing tool selection; it does not establish an ideal tool count. Clash: remove superseded instructions (06:55) Correcting the retry mistake leaves both the original note and its correction in the history. The model now sees incompatible accounts of the same behavior. Remove the obsolete claim from active context, or start a fresh window with the corrected task state. Appending a correction is not always enough. “An agent can get better at long tasks, by holding less.” — Martin Keen, 09:54 ...