Machine-native execution routing for AI agents

One supported task in. A result back.

ATL routes supported search, extract, summarize, translate, and classify tasks to eligible providers, executes through the selected route when ATL owns execution, applies bounded fallback when appropriate, and records durable outcomes.

Connect ATLTry the routing demoSee measured production results
1. Describe the taskNo provider identity required.
2. ATL checks eligibilityHealth, freshness, policy, budget and Gate.
3. Execute with bounded fallbackFor ATL-owned execution, the selected route runs and may fail over within policy.
4. Return result + OutcomeThe real execution result and durable Outcome close the loop.
Measured on the public production path · 2026-09-25

Less routing overhead. Smaller machine responses. Fail-closed behavior preserved.

After the first Decision hot-path performance pass, ATL was benchmarked from an external GitHub-hosted runner against the public production endpoint. These are measured results from that test set, not theoretical projections.

~1 msInternal Decision latency observed in the production comparison request.
1.75 sMedian external round trip from the GitHub runner to the public ATL endpoint and back.
91.3%Smaller serialized Decision response body with compact mode: 9,033 B → 783 B.
16 / 1612 supported routes returned the expected primary + fallback; 4 unsupported capabilities returned NO_MATCH.

Scope matters: the 91.3% figure is response-body reduction for this benchmark, not a claim about model tokens or total application bandwidth. The 16/16 figure describes this test set, not universal accuracy. Read the benchmark methodology and caveats.

Live routing demo · Decision only

This browser demo intentionally shows a real routing Decision only. For end-to-end execution of supported tasks, connect ATL and use atl_complete_task; ATL then owns route selection, execution, bounded fallback when appropriate, and durable Outcome recording.

Task
Ready.
Routing Decision
Run the routing demo to see the selected provider, alternatives and decision reference.

Connect where you already build

MCP

https://mcp.agenttrafficlab.com/mcp

Use atl_complete_task by default for supported tasks. Use atl_decide only when you specifically need routing without immediate ATL-owned execution.

Python

from agenttrafficlab import decide\n\nroute = decide("summarize this ticket")

Routing client · zero runtime dependencies.

JavaScript

import { decide } from "@agenttrafficlab/client";\n\nconst route = await decide("summarize this ticket");

Routing client · Node.js 18+.

Why ATL exists

Not just model aggregation

ATL is focused on route eligibility, lifecycle state, Gate checks, authorization, bounded failover and outcome evidence.

Fail closed before routing

A provider is not treated as production-eligible just because an endpoint exists.

Machine traffic first

The target is agent-to-infrastructure traffic, not a dashboard that humans must operate manually.