Skip to content
Corvid Docs

Performance

What’s slow, what’s fast

Corvid’s optimization target is the AI-orchestration surface, not raw throughput. Honest numbers (Phase 17 baselines):

  • Orchestration overhead (no model call): ~25–36× slower than Python LangChain on equivalent code paths. ~1.7–2.6× slower than TypeScript on refund_bot, rag_qa_bot, support_escalation_bot.
  • Compile-time guarantee surface: zero runtime cost. approve, Grounded<T>, and effect rows are typecheck-time constructs.
  • Replay: byte-identical re-execution at memory speed (no network).
  • Codegen: Cranelift backend; comparable to debug-mode rustc for agent-shaped programs.

The “honest slower” framing is deliberate — Corvid’s value is the correctness surface, not raw orchestration speed. See the benchmarks page for full archives.

When to optimize

In order:

  1. The model. Most of the latency and cost is the LLM call. Provider routing (requires clauses) and prompt tightening dominate.
  2. The retrieval. RAG quality > LLM quality > orchestration speed.
  3. The agent shape. Agents that loop over n calls cost n × per- call cost; loop bounds and budget annotations are cheap to add.
  4. The orchestration. If after the above three you’re still bottlenecked on Corvid’s orchestration overhead, drop the inner loop into Rust via the FFI.

Budgets are real

@budget($X) is a compile-time ceiling. The compiler computes the worst-case composed cost; the binary refuses to ship if it exceeds X. Use this to keep production agents from drifting upward in cost as they accrete features.

Concurrent jobs

The durable runner (corvid jobs run --workers=N) scales to N async workers per process. Past a few thousand jobs/second per host, scale horizontally — multiple runner processes against one queue.

Profiling

Terminal window
corvid run src/main.cor --profile

Emits a flamegraph for the run. Useful for finding hot paths in agent composition (rare) or in custom Rust FFI calls.

Reproducible benchmarks

Terminal window
corvid bench src/main.cor
corvid bench compare python --source src/main.cor target/python_traces

Published baselines live in benches/ and ship via Phase 22’s reproducibility scripts.