Skip to content

Instantly share code, notes, and snippets.

@jwmatthews
Created May 12, 2026 20:43
Show Gist options
  • Select an option

  • Save jwmatthews/221c436ff36e07e80e129d7cf4b91c45 to your computer and use it in GitHub Desktop.

Select an option

Save jwmatthews/221c436ff36e07e80e129d7cf4b91c45 to your computer and use it in GitHub Desktop.
Dex Horthy: Key Findings on Agentic & Context Engineering

Dex Horthy: Key Findings on Agentic & Context Engineering

The Central Thesis: Context is the only control surface

LLMs are stateless and non-deterministic. Every decision an agent makes is determined entirely by the tokens currently in its context window. The implication is stark: better tokens in, better tokens out. And critically, more tokens does not mean better outcomes. DEV Community

The "Dumb Zone"

Performance degrades beyond roughly 40% context utilization. As context usage grows, model quality degrades, often beginning around 40% of the window depending on task complexity. Common culprits: large tool outputs (JSON blobs, UUIDs, logs), unfiltered file dumps, long correction-loop histories, and MCPs dumping irrelevant data. Once you're in the dumb zone, model quality doesn't matter. Alexlavaee | DEV Community

Trajectory Poisoning

If a conversation's pattern is mistake, correction, mistake, correction, the most likely continuation is another mistake. Bad trajectories reinforce failure modes. This is why restarting sessions or compressing context often beats continued correction. DEV Community

Intentional Compaction

This is the core practice. The philosophy is to spend as much time as possible on intentional context management. Instead of dragging an ever-growing conversation forward, you summarize the current state into a compact markdown artifact, validate it as a human, and start a fresh context seeded with that artifact. You compact relevant files, verified architectural behavior, decisions made, and constraints. You strip raw logs, tool traces, and repetitive error chains. Compaction converts exploration into a one-time cost instead of a recurring tax. Medium | LinearB

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment