Agent patterns
Multi-agent compression
In multi-agent systems, every agent sees the messages from all other agents. With 5 agents exchanging 3 messages each, that's 15 messages per task — 3,000-6,000 tokens of inter-agent chatter.
The communication graph
Each agent receives: its own instructions (500-1,000 tokens), messages from other agents (200-500 tokens each), shared context (1,000-2,000 tokens), and tool call results (500-1,000 tokens). For a 5-agent team, a single task can consume 5,000-10,000 tokens before any agent generates a response.
Frequently asked questions
Which messages should I compress?
Focus on agent-to-agent messages. System instructions and tool schemas should be preserved.
Does compression break agent coordination?
No. Only irrelevant or redundant messages are removed. Coordination-critical messages are preserved.
Build with less context
Put compression in front of your next LLM call.
Use the hosted API or run SuperCompress locally. Keep the evidence, drop the token waste, and measure the savings before it reaches your model.