Cendriix Orchestrated is not another LLM. It is a routing-and-ensemble layer that, for each step of your run, selects the best backbone from Claude, GPT, Gemini, DeepSeek, and others, then runs a critic pass and falls back across providers on failure. The result: lower cost than always-Opus, higher reliability than any single model, and resilience to provider outages your agents never have to handle themselves.
The routing decision happens at step granularity. A single agent run may use Claude for its planning step, DeepSeek for the code diff, Haiku for the PR summary, and Gemini for a visual artifact, all without any configuration on your part.
Active route: Claude Opus, reasoning-heavy steps. Matches model-router + Bedrock adapter in cendriixai-backend; not a static sequence diagram.
Reasoning-heavy steps, Claude Opus
Tool use / JSON, GPT-5
Code diffs, DeepSeek V3
Extraction / vision, Gemini
Summaries / triage, Haiku / Flash
Why not just pick one model
Six reasons single-model pipelines break in production
Provider outages happen
Every major LLM API has logged multi-hour outages in production. A single-model dependency means your entire workflow stops. Cendriix Orchestrated detects degraded providers and re-routes in under 200 ms.
Opus pricing for boilerplate is waste
Summarising a PR description or extracting a JSON field does not need $15/1M-token reasoning. Cendriix Orchestrated sends those steps to a fast, cheap backbone and reserves frontier spend for tasks that justify it.
Task-fit matters more than raw benchmark score
Structured extraction favours Gemini. Multi-file code diffs favour DeepSeek. Deep security reasoning favours Claude. No single model wins every category, Cendriix routes each step to the model best matched to that step type.
Quality regressions go undetected
Model providers ship silent updates. A backbone that scored 94 % on your eval last month may score 81 % today. Cendriix runs a continuous 7-day eval window per backbone and demotes models that regress below threshold.
Rate limits hit at the worst moment
A burst of concurrent runs can exhaust a single provider's TPM quota mid-pipeline. Cendriix Orchestrated distributes load across providers and holds a credit reserve on each, so rate-limit hits stay invisible to your agents.
Context windows are a moving target
A 64 K-token model silently truncates a 90 K context, producing subtly wrong output. Cendriix measures the context for each step and selects a backbone whose window fits, without you having to audit every prompt.
What it does per step
Six decisions, sub-second, invisible to your agent
1
Classify the step
A lightweight classifier identifies the step type, reasoning, code generation, extraction, summarisation, or tool orchestration, and tags it with a confidence score.
2
Pick the backbone(s)
The router consults live cost, latency, and quality signals for each provider. It selects one primary backbone and, for high-stakes steps, an optional secondary backbone for ensemble reconciliation.
3
Optionally ensemble and reconcile
For steps classified as high-consequence, a policy decision, a prod-deploy plan, a security analysis, Cendriix runs two backbones in parallel and applies a fast reconciliation pass to produce a single output.
4
Run the critic pass
A compact critic model verifies the output against the step's input and expected output schema. If the critic scores below threshold, the step re-runs on the next-best backbone automatically.
5
Fall back on failure
If a provider returns an error, times out, or the critic rejects the output after one retry, the orchestrator falls back to the next-ranked backbone with no pause visible to your agent.
6
Log every choice to the audit trail
Every routing decision, which backbone was picked, why, what the critic scored, and whether a fallback occurred, is appended to the run's immutable audit log. View the audit log format.
Tuned on your workspace
The more you run, the better it routes.
Cendriix Orchestrated maintains a 7-day rolling evaluation window for each backbone on each step type your workspace exercises. When a model's observed quality drops below its baseline, even if the provider has not announced any change, the orchestrator demotes that backbone automatically and logs the reason.
If a subsequent eval window shows the model has recovered, it is promoted back without any action from you. The entire adapt-and-rollback cycle is automatic.
# routing-outcomes.log (excerpt)
code.diff
winner: deepseek-v3 (0.94)
alt: claude-opus: 0.91
plan.multi_step
winner: claude-opus (0.97)
alt: gpt-5: 0.89
extract.json
winner: gemini-flash (0.98)
alt: haiku: 0.95
summarise.pr
winner: claude-haiku (0.93)
alt: gpt-mini: 0.90
Cendriix never trains on your source code, secrets, or task payloads. Adaptation is based exclusively on routing outcomes, which backbone produced the highest-scoring output for a given step classification, not on the content of your runs.
Still fully model-agnostic
Opt-in by default. Never a lock-in.
Cendriix Orchestrated is the recommended default, but you can pin any backbone on any step or for an entire agent at any time. Pin Claude Opus everywhere, pin DeepSeek for cost-sensitive batch runs, or mix, the YAML is the source of truth and you own it.
Pin a model globally
model:
default: claude-opus-4-7
One line in the agent YAML overrides the router for every step.
Cendriix Orchestrated is not a premium add-on. It is the default mode for all agents on every plan tier. You are billed at blended backbone cost, typically 40–70 % less than pinning Claude Opus exclusively, with no markup on the routing layer itself.