How do you handle context degradation when chaining multiple tools in an AI Agent?

As tool ecosystems grow and agents call multiple external tools in sequence, prompt bloat and verbose tool returns start degrading context rapidly.

In your workflows, do you prefer:

1. Truncating or summarizing previous tool outputs before sending the next turn?
2. Spawning isolated subagents and passing only synthesized state back to the coordinator?
3. Enforcing strict, minimal schemas to keep payload footprints small?

Would love to hear what patterns have proven most reliable in your production setups!