Why AgentFlight & Verified Proof
Multi-step tool-calling agents fail in ways standard HTTP mocks cannot reproduce cleanly: key-order drift across runs, silent parameter omissions on weak MCP schemas, A↔B tool oscillation loops, and $5 overnight token spikes in CI.
Created independently by Akandwanaho Ivan Alvin, AgentFlight records every step into an atomic SHA-256 cassette with API keys scrubbed automatically so teams test deterministically offline at $0.00 spend.
Before vs. After Schema Hardening (38 → 100)
Inspect how agentflight fix and ToolSmellLinter eliminate MCP schema failures in a single deterministic pass.
TS001 (-12)Tool description under 15 characters left model routing ambiguous.TS005 (-5)Boolean flagmodehad unclear polarity without explicit description.TS006 (-6)Unconstrainedstatusstring lacked an explicitenumlist.TS010 (-10)Destructivedelete_recordlacked aconfirmordry_runguard.
Measured Subsystem Benchmarks (Python 3.12)
Reproducible directly on any machine with agentflight benchmark:
| Subsystem Operation | Workload | Total Time | Per Unit | Throughput |
|---|---|---|---|---|
FlightRecorder.record (SHA-256 chain) |
1,000 spans | 7.42 ms | 7.42 us | 134,770 spans/s |
verify_cassette (tamper check) |
1,000 spans | 3.40 ms | 3.40 us | 294,117 spans/s |
CassetteReplayer.replay_step ($0 stub) |
1,000 spans | 2.83 ms | 2.83 us | 353,356 spans/s |
TrajectoryCircuitBreaker.observe_span |
1,000 spans | 3.01 ms | 3.01 us | 332,225 spans/s |
ToolSmellLinter.lint (TS001–TS010) |
200 tools | 119.27 ms | 596.35 us | 1,676 tools/s |
How AgentFlight Compares
| Capability | AgentFlight | Live LLM CI Runs | Hand-Rolled Mocks |
|---|---|---|---|
| CI run cost & latency | < 3 us/span · $0.00 | $0.05–$5.00 · 10–45s | Free but fragile |
| Cryptographic tamper protection | Canonical SHA-256 chain | None | Silent fixture drift |
| Automatic API key scrubbing | Enabled by default | Manual env sanitizers | Manual redaction |
| Ping-pong loop & cost circuit breaker | Built-in pre-flight guard | Stops after budget alert | Not included |
| MCP schema linting & auto-repair | TS001–TS010 + fix CLI | Manual prompt tuning | Not included |
Get Started
Read the Documentation, audit schemas live in the Interactive Studio, or browse all 15 linter & guard rules.