Module 12 · cognitive feature

Active intent routing

Not every query deserves a frontier model. A low-latency classifier reads the prompt and hands it to the specialist that owns that lane — cutting inference cost, cutting latency, and cutting hallucinations by keeping each agent inside its own tool set. When the classifier is not confident, it escalates rather than guesses.

$ python -m fde_toolkit router --query "..."

Route a query

routedSQL / Data Agentworkhorse-8bconfidence 1.00

Matched: exposure, total

tools in lane: query_warehouse, describe_entity

SQL
2
Compliance
0
General
0
Investigation
0
Ops
0

Specialist roster

SQL / Data Agentworkhorse-8b
Reads the semantic layer and answers quantitative questions.
$0.00080/1k · 320ms · hallucination 6.1%
Compliance Agentreasoner-midguardrailed
Interprets policy, regulation and internal control documents.
$0.00900/1k · 950ms · hallucination 2.8%
Investigation Agentfrontier-xlguardrailed
Multi-step alert triage that must weigh conflicting evidence.
$0.04500/1k · 2400ms · hallucination 1.7%
Ops / Integration Agentworkhorse-8b
Runs the plumbing: job status, extracts, connector health.
$0.00080/1k · 320ms · hallucination 6.1%
General Assistantworkhorse-8b
Fallback for chit-chat and open questions with no enterprise scope.
$0.00080/1k · 320ms · hallucination 6.1%

This query

600 output tokens

Routed: $0.00060
Frontier for everything: $0.02700
Difference: −97.8%
Latency: 365ms vs 2400ms

Batch of 5 typical queries

What is our total exposure to NMC Health this quarter?
SQL / Data Agent$0.00060 vs $0.02700
Is this transfer to an Iranian beneficiary permitted under our AML policy?
Compliance Agent$0.00552 vs $0.02700
Investigate alert AML-4471 and tell me whether to escalate it.
Investigation Agent$0.02712 vs $0.02700
The nightly KYC extract job failed again — what is the status?
Ops / Integration Agent$0.00060 vs $0.02700
Can you look at this?
General Assistant$0.00060 vs $0.02700

Routed total$0.03444

Frontier-only total$0.13500

Saved on the mix74.5%