We built FlashCompact because context compaction is one of the most broken parts of modern coding agents.

Right now, compaction is:

• slow (often 2+ minutes)
• lossy (kills useful context)
• expensive (wastes tokens and steps)

So people avoid it… and agents degrade over time.

What we built

A specialized model for context compaction:

• 33,000 tokens/sec throughput
• 200k → 50k tokens in ~1.5s
• Designed specifically for agent workflows, not general summarization

Compaction should feel invisible. This actually does.

Why this matters

We analyzed 200+ real agent sessions across 40+ top coding agent harnesses.

The core insight:

Most context bloat comes from tool responses, not model generation.

That means naive summarization fails, because it compresses the wrong things.

FlashCompact:

• removes redundant tool output
• preserves execution-relevant state
• reduces context without hurting task performance

Result:

• no performance drop
• fewer tokens
• fewer steps

The deeper point

Agents don’t fail because models are weak.

They fail because their context degrades over time.

If you want agents to:

• run longer
• take more steps
• stay coherent across complex tasks

You need clean, compressed context.

https://docs.morphllm.com/sdk/components/compact