J2Platform Intelligence
LLM Model Routing & Ops Agent
Chooses which underlying model (Claude, GPT, Gemini, or an open-weight model) each agent task should call, based on cost, latency, and quality targets, and manages fallback and retries across providers.
How it runs
- What it reads
- Task type, historical per-model quality scores, live provider latency/cost signals, per-project budget.
- What it produces
- Routing decision log, provider fallback events, model performance leaderboard.
- When it runs
- Continuous (every LLM call made by any agent).
The same guardrails as every other agent
LLM Model Routing & Ops Agent is governed exactly like the rest of the fleet. Its rule file is versioned, never edited in place. Its tool list is a declaration, not a grant — it can only call a tool a gateway policy has explicitly allowed in that project. Every tool call it makes is policy-checked and written to the audit log before it leaves the platform, and a project's cost cap is checked before a run starts, not reconciled after it.
More in Platform Intelligence
- J1Knowledge Base / RAG Grounding AgentIngests and indexes the project's codebase, PRDs, ADRs, prior cycle outputs, and uploaded reference material into a governed vector store, and serves grounded retrieval to every other agent so outputs stay consistent across a 44-agent fleet and hallucination is reduced.Read the product page
- J3Platform AI Safety & Red-Team AgentContinuously adversarially tests XOOZLY's own agent fleet — prompt injection, tool misuse, authority creep, jailbreak attempts — distinct from the Security & Penetration Testing Agent, which tests the software the fleet builds, not the fleet itself.Read the product page
- J4Agent Marketplace & Template Curator AgentCurates, versions, and publishes the Agent template library — vetting community or internal templates for quality and security before they become installable by a project.Read the product page
- J5MLOps AgentOwns the training, evaluation, and drift-monitoring lifecycle for any custom ML model ML Feature Engineering Agent builds — champion/challenger comparison, automated retraining triggers, and production model-quality gates; distinct from Agent Quality & Evaluation Agent, which scores the fleet's own agent output, not a trained model's prediction quality, and from LLM Model Routing & Ops Agent, which routes calls to third-party foundation models rather than operating models this platform trains itself.Read the product page
- J6LLMOps AgentOwns prompt-version management and evaluation for every system prompt and tool definition the agent fleet uses — versioning, A/B testing a challenger against the production default, LLM-as-judge quality/faithfulness scoring, and hallucination-rate regression detection; distinct from LLM Model Routing & Ops Agent, which chooses which third-party model serves a call, not what prompt that call sends, and from MLOps Agent, which operates custom-trained ML models this platform builds itself, not the fleet's own prompt text.Read the product page