01
Reuse stable prompts
Provider cache controls reduce the cost of repeatedly sending long system instructions, tool schemas, and stable project context.
// Performance
Faster turns. Cheaper tokens.
Agent performance is more than model speed. PasClaw reduces repeated prompt work, overlaps independent tools, and keeps retrieval close to the process.
Prompt cachingParallel toolsHybrid memoryLocal retrieval
02
Dispatch independent tools
Multiple tool calls from one turn can run concurrently, cutting wall-clock time for network-heavy or multi-source tasks.
03
Combine keyword and semantic recall
SQLite full-text search handles exact terms while local embeddings recover conceptually related memories without an external vector service.
04
Measure the whole loop
Native startup, provider latency, tool execution, retrieval, and token usage all contribute to the experience and can be optimized independently.
From the documentation
Technical details
Implementation notes drawn from the PasClaw repository documentation.
Reduce serial work
PasClaw overlaps independent read-only tools and uses execute_code to collapse multi-step local processing into fewer model round trips.
- The repository reports roughly 50% lower wall time for representative multi-network-tool turns.
- Its tiered fs_grep implementation is reported at 3–10× the earlier baseline.
- These are repository engineering estimates, not independently verified benchmark results.
Measure cost and latency
Prompt caching, output truncation, local retrieval, and observability address different parts of total turn time.
- Cache read/write counters expose provider reuse.
- tool_output_cap limits oversized observations before the next inference.
- The stats endpoint and OpenTelemetry spans separate model, tool, and orchestration costs.
Source documentation
Common questions
Frequently asked questions
Keep exploring
See how the rest of PasClaw fits together.