crewai - 💡(How to fix) Fix [FEATURE] Token-efficient serialization for agent communication to reduce context overhead at scale [1 participants]

Feature Area

Agent capabilities

Is your feature request related to a an existing bug? Please link it here.

NA

Describe the solution you'd like

Native support for pluggable serialization in the Crew execution pipeline, with ULMEN as the first alternative to JSON.

ULMEN is a serialization engine built specifically for LLM agent pipelines.

Benchmarks on NVIDIA Tesla T4 production hardware:

• 44% fewer tokens vs JSON
• 3x faster deserialization vs orjson
• 4.1x smaller wire format
• $59K saved per 10M agent loops on GPT-4o

Beyond compression, a Semantic Firewall validates agent state transitions before they reach the LLM:

• Rejects orphaned tool calls
• Catches invalid step transitions
• Structured errors instead of silent failures

Implementation would be fully opt-in:

crew = Crew( agents=[researcher, writer], tasks=[research_task, write_task], process=Process.sequential, serializer="ulmen" )

Drop-in Python/Rust. No schema compilation. Pure Python fallback if Rust unavailable.

Reproducible benchmark notebook: github.com/makroumi/ulmen

Describe alternatives you've considered

• orjson: faster than json but identical token count. Doesn't address context window overhead.

• MessagePack: reduces wire size but not token count. No semantic validation layer.

• Manual prompt compression: lossy, requires custom logic per use case, not systematic.

ULMEN is the only approach that simultaneously addresses token count, wire size, speed, AND semantic validation in one drop-in library.

Additional context

CrewAI's sequential and hierarchical process patterns are particularly exposed to this problem because:

1. Each agent receives full task context as JSON
2. Tool results serialize back as JSON each turn
3. Key names repeat identically across every turn
4. A 5-agent crew compounds the overhead 5x

At 10M crew loops the JSON tax is ~$59K on GPT-4o pricing. Not a rounding error.

Silent tool failures passing through JSON undetected are also a leading cause of crew

<img width="1861" height="1281" alt="Image" src="https://github.com/user-attachments/assets/6a83fbd8-eb74-4a0c-a9e0-ff4d9dbab313" />

hallucinations that are difficult to diagnose.

Live benchmark notebook to verify numbers: github.com/makroumi/ulmen

Willingness to Contribute

Yes, I'd be happy to submit a pull request