Tool · Encoder
English → AACP packet
Real examples. Loss flags are explicit, not every prompt compresses cleanly.
FETCH|HR|return:HR-Agent|p:1|aacp:1.4|res:emp_salary|period:2024-08|filter:status=active|fmt:json
Compression status · NO LOSS
All structural intent preserved. Safe for agent-to-agent routing.
Token counts for the HR example are measured from live API usage_metadata ( → tokens, -7.1%). Other examples use estimated counts. Measured benchmarks for the full payroll workflow are on the home page.
v1.4 · Rule-based encoder
Zero LLM calls. Deterministic output.
For known workflow types, a rule-based encoder generates packets at zero LLM cost. Zero API cost. Same input always produces the same packet.
For novel instructions outside known workflows, FallbackEncoder routes to the LLM and logs the result to a local registry as a future rule-based candidate. One LLM call per novel pattern, reused forever after.
| Metric | LLM encoder | Rule-based |
|---|---|---|
| Coordination tokens (measured) | ~-23% | ~-23% |
| Encoder LLM calls | 1 / instruction | 0 (known workflows) |
| Encoder cost | ~$0.0006 / call | $0.00 |
| Novel instructions | LLM encoded | Fallback + logged |