Harvey and Baseten Deep Dive: Building Reliable Long-Horizon Agents with 100M-Token Context and KV Cache Compaction
baseten · x · 2026-07-23
Harvey and Baseten co-founder Gabe Pereyra and model training leads discuss the core challenges of building agents that can reliably complete tasks over hours or days.
Key topics:
- 100M-token data rooms for M&A agents handling ultra-long contexts.
- Neural KV cache compaction to overcome search and reasoning bottlenecks in long contexts.
- Synthetic client matters for model training data generation.
- Continual learning to adapt agents over time.
The conversation also covers current limitations of agents in search and long-context windows, and how techniques like KV-cache compaction, synthetic data, and continual learning could help.
Related event: Harvey and Baseten Discuss Challenges of Building Long-Running Agents(4 posts)→
More from coding & agent
- LLMs turn old-school Burp Spider workflows into a mess of prod-db mistakes — rez0__ · 2026-07-23
- Claude set up an Unreal Engine MCP while the user went for ice cream — stspanho · 2026-07-23
- Switchloom open-sources deterministic model routing for coding agents — kevinkern · 2026-07-23
- A personal AI home setup uses Tailscale, Termius and an always-on Mac — daniel_mac8 · 2026-07-23
- Claude Code one-shot config turns a Corsair keyboard into an agent control deck — SimonBalmain · 2026-07-23
- Most “self-improving” AI agents don’t improve without real-world verifiers — HamelHusain · 2026-07-23