Agent roster

Mirrors mesh_agent_registry.py, capability-routed and cost-governed

replay snapshotrecorded 2026-07-11T17:35:27Z7 agents
Alex Seller Botactive
lead_qualificationconversation_analysis
34 tasks today1067ms avg$0.41 cost
Lead Lifecycle Botactive
followup_automationvoice_interaction
58 tasks today423ms avg$0.22 cost
Intent Decoderactive
conversation_analysis
92 tasks today187ms avg$0.17 cost
Conversation Intelligence Serviceidle
conversation_analysismarket_intelligence
21 tasks today512ms avg$0.09 cost
Enhanced Property Matcheractive
property_matchingmarket_intelligence
17 tasks today1490ms avg$0.13 cost
Ghost Followup Engineidle
followup_automation
44 tasks today95ms avg$0.04 cost
ML Analytics Engineidle
conversation_analysismarket_intelligence
12 tasks today233ms avg$0.02 cost

Local recording: cache artifact benchmarks/results/cache_local_20260711T173515Z.json is passed. RedisClient completed connect, health check, set, get, TTL, delete, and disconnect against local Redis. Mesh artifact benchmarks/results/mesh_local_20260711T173515Z.json is passed. Two local generic agents completed 200 tasks, exercised budget validation, and entered maintenance after emergency shutdown. The roster and governance values below remain canned replay data, not live mesh telemetry.

3-tier response cache

claude_orchestrator.py, exact, normalized, semantic

L1In-memory exact match
91% hit3msmeasured
L2Redis normalized prompt
62% hit12msmodeled
L3Redis semantic similarity
31% hit26msmodeled

L1 hit rate measured live (seed 42, benchmarks/results/cache_live_2026-05-27.json). L2/L3 are modeled targets, labeled to keep the numbers honest.

Cost governance

agent_mesh_coordinator.py, budget gates and shutdown

$1.08spend this hour of $50/hr budget
$0budget $50shutdown $100
Tokens saved by cache today84,312

Every task is admitted through a budget check before routing; the coordinator records per-agent cost and shuts the mesh down at the hard ceiling. Audit trail in docs/adr/0004 and 0011.