CONTROLLED COMPARISON

Route the same agent task two ways.

Observe the tradeoff. Keep the claims smaller than the evidence.

Scenario Developer question Controlled context

Model routes

A Economy route

Planner Answer

B Capability route

Planner Answer

Evaluator

Evidence probes Repeat stable prefixRecord returned cache evidence—or its absence. Inject recoverable errorCapture an invalid-model failure, then continue.

Waiting for live model discovery.

LOCAL SAFETY

The key stays server-side. Live execution is bound to localhost. Saved traces contain evidence, never credentials.

EVIDENCE LEDGER

No comparison recorded

Load the sample to inspect the interface, or connect locally for a live run.

Fastest observed—

Cheapest observed—

Evaluator—

Observed cost—

WHAT THIS RUN MEASURES Two model routes for the same diagnostic prompt

The prompt describes a suspected context slowdown. This trace does not measure a before/after context latency baseline; that requires a separate repeated experiment with only context size changed.

This surface will show two model routes, the MCP boundary, streaming first-token time, returned token usage, calculated cost, retry history, and one deliberately limited evaluation.

Open-stack extension

LangGraph → Tavily → Token Factory

A separate three-node LangGraph example performs agentic search with the official Tavily SDK, selects evidence, then passes bounded sources to a Token Factory-compatible answer step. One keyless Tavily search was captured live; its final summary was deliberately local, so no additional model tokens were spent.

Inspect graph source ↗ Open live search trace ↗

The cheaper route lost the evaluation.

No comparison recorded

Open-stack extension