Invited Talks
Proactive Context Management for Long-Horizon Code Agents
TU Munich, Advanced NLP Masters’ course · Guest lecture · June 6, 2026
Gave a guest lecture on one of my passion topics: proactive context management for long-horizon code agents. As code agents operate over longer horizons, context management has evolved from an afterthought to a first-class post-training topic.
- Until recently: the harness manages context by prompting the model to summarize once it crosses a heuristic threshold (e.g., 80% of the context window).
- Now: learned meta-actions, where the model itself recognizes when a tool output is stale, or a subtask’s branch is dead, and folds it away on its own initiative.
We closed the session with a hands-on demo of using subagents to scale qualitative trajectory analysis, automating the painful part of debugging why an agent went sideways over a 100+ step run.
Thanks to Alexander F. and Miriam Anschütz for making this possible.
Context Engineering for Agentic Hybrid Applications
Berlin DataTalksClub · Invited talk with Ivan Potapov · March 10, 2026
Discussed the cost and reliability implications of context growth in long-horizon production agents, arguing for efficiency-focused, proactive context management. Built on findings from The Complexity Trap that simple observation masking matches LLM summarization at a fraction of the cost.