Engineering Discussion: Optimizing Context Compaction in DeepSeek Harness
shady101852 · reddit · 2026-08-28
A developer shared engineering experiences using the open-source DeepSeek Harness. Having fixed bugs in the compaction feature, they found the context still consumes 130k-150k tokens (max 500k) post-compaction and aim to optimize it down to 50k-90k to reduce costs. The post sparked technical discussion on long context management strategies.
More from coding & agent
- One agent sitting in on all your meetings, queryable by everyone, gets you close to total info transparency — rohanpaul_ai · 2026-08-28
- The biggest mistake companies make with agents: one generalist trying to do everything — annetgriffin · 2026-08-28
- Agent irreversible action failure costs more to undo manually — AivenAdmin1 · 2026-08-28
- Perplexity Launches Portable Computer: Fully Local, Cloud-Free AI Agent — AravSrinivas · 2026-08-28
- LangChain releases free course on agent engineering fundamentals — LangChain · 2026-08-28
- Review: open-source desktop app for reviewing AI-generated code — ycombinator · 2026-08-28