Showcase agents / Persistent Agent
Long-running agent builders
Persistent Agent
What problem? Multi-day agent goals lose state on crash — operators need checkpoint visibility, HITL waits, and session-linked traces.
Durable execution showcase: checkpoints, episodic memory, HITL waits, cron wake, and crash resume across sessions.
What Inferix observes
Every run dual-writes to LensAI and TraceForge. Switch to this tenant in the console to see live KPIs.
- TraceForge — long-lived trace linking across sessions
- LensAI — events tagged with checkpoint_id and session_ordinal
- RouteIQ — sticky model routing per checkpoint
- DriftWatch — goal-completion quality over multi-day runs
Tools & models
Real SQLite backends, not mocked HTTP stubs.
- state.load_checkpoint
- state.save_checkpoint
- memory.retrieve
- memory.write
- workflow.resume
- hitl.request_approval
- scheduler.wake
- domain.act
Model routing
- Sticky RouteIQ across sessions
- cheap summarize for memory
- escalate for domain.act
Try it
Start the stack, then run the golden scenario. Traffic appears in the console within seconds.
1. Start stack + agents
cd inferix && ./run-inferix.sh --full --agents
2. Golden scenario
curl -s -X POST http://localhost:9107/scenarios/resume_after_crash | jq
Expect HTTP 200 with JSON status and tool steps. Traces appear under this tenant in ~10s.
3. All agents at once
cd inferix && ./scripts/generate-agent-traffic.sh
Expected: JSON with status and tool steps. Then open the tenant console and filter traces by tenant-persist. In-browser runner stays deferred — curl is the verified path.
Console: /console/overview?tenant_id=tenant-persist · API traces: GET /v1/traces?tenant_id=tenant-persist
See all seven agents in the console
Multi-tenant super-dashboard. Filter by tenant, drill into traces and costs.