Seven Frameworks.
One AI Control Plane.
Framework Ecosystem
Seven Frameworks.One AI Control Plane.
CARDIAC-PURR AI Control Plane integrates with seven agent frameworks and routes each request through the same capability-aware model qualification and tier selection used across the platform. Agent workloads route through the same governed, capability-aware layer as every other request — no separate integration tier, no separate rules. This page covers the architecture and integration paths — for the full AI Control Plane provider benchmark, see Benchmarks.
What is CARDIAC-PURR AI Control Plane? CARDIAC-PURR AI Control Plane is a session-aware AI routing and governance layer that automatically selects the lowest-cost capable model for each request across seven agent frameworks and nine providers. On a 100-query technical benchmark (September 2026, build V120, cache bypassed), it achieved 91.2% lower observed model inference cost versus a realized LARGE-tier reference arm, with 93 of 100 requests correctly routed to the small tier, 4 to medium, and 3 to large — 0 cascades, 0 retries.
LangChain
Routed through the standard CARDIAC-PURR AI Control Plane model interface.
LlamaIndex
Integrated via its ReAct and FunctionTool agent interfaces — the most specifically-documented integration path of the seven.
AutoGen
Multi-agent conversations routed per request through the same capability-aware layer.
CrewAI
Crew and task-based agent workflows route through the control plane per step.
Haystack
Pipeline-based agent components route through the same model qualification layer.
DSPy
Programmatic prompting pipelines route through CARDIAC-PURR AI Control Plane's tier selection.
Semantic Kernel
Plugin and planner-based agent execution routes through the control plane.
Compatible integration path
An OpenAI-compatible endpoint is available for applicable frameworks — point your existing client at the CARDIAC-PURR AI Control Plane base URL without a framework-specific adapter.
Architecture
The AgenticControl Plane.
Routing
Automatic.Or Explicit.
Live tier-resolution trace
A current v114 validation call using explicit LARGE-tier routing:
Reflects current production configuration (build cardiac-purr:v114). This is a capability trace, not a benchmark result, and is separate from the dated benchmark evidence below.
What "auto" does today
model="auto" evaluates the request and selects a tier without manual specification. In the published agentic benchmark below, this resulted in every one of the 350 factual-QA requests routing to the small tier — consistent with a workload that didn't require medium- or large-tier capability.
Recorded Benchmark Evidence
Two Runs.Different Questions. Both Published.
Experiment 1 — Agentic Framework Routing
What this benchmark does & doesn’t show
- 01These figures apply only to the published 350-request factual-QA workload — they are not a general accuracy or savings claim.
- 02Every request in this benchmark routed to the small tier; medium and large tiers were not exercised by this run.
- 03Caching affected the workload. This benchmark reflects a cached-influenced result, not a purely cold-inference measurement.
- 04Results are reported as an aggregate across all 7 frameworks combined — no individual per-framework breakdown is published from this run.
Experiment 2 — Economic Cost Benchmark (Corrected Methodology)
What this benchmark does & doesn’t show
- 0191.2% is the observed cost reduction vs a realized LARGE-tier reference arm — the models that actually ran, not a theoretical always-LARGE price. The 66.6% figure uses a synthetic always-LARGE baseline; both are published because they measure different things.
- 020 cascades and 0 retries were recorded. This was a successful-routing economic test — failure-path and escalation economics were not exercised.
- 03Answer quality was not measured on this cohort. Separate prior evaluations produced results in the 96–100% range on different datasets; quality and cost have not yet been co-measured on the same requests.
- 04Routing overhead (latency cost) was not separately measured in this benchmark. Full methodology including prior experimental runs is documented in the published DEV.to article.
Per-Tier Cost Breakdown
| Tier | Queries | CP Cost | LARGE Ref Cost | Reduction |
|---|---|---|---|---|
| SMALL | 93 | $0.163543 | $3.715975 | 95.6% |
| MEDIUM | 4 | $0.065385 | $0.156775 | 58.3% |
| LARGE | 3 | $0.121770 | $0.125170 | 2.7% |
| Total | 100 | $0.350698 | $3.997920 | 91.2% |
| Avg per query: CP $0.003507 · LARGE ref $0.039979 | ||||
The 2.7% reduction on LARGE-tier requests confirms the router sent those requests to the large tier because they required it — savings came from precision routing, not indiscriminate downgrading.
Evidence Boundaries
What This PageDoes Not Claim.
Structured-output, reasoning, context-window, and vision awareness are tracked as capability metadata dimensions within the routing layer. No benchmark results are currently published for these specific dimensions — only the factual-QA (Experiment 1) and economic cost (Experiment 2) workloads on this page have been benchmarked.
Streaming is supported in current integration code. No streaming-specific benchmark results are published on this page.
No per-provider capability matrix with scored values is published. The architecture diagram above reflects qualification stages that exist in the routing layer, not a comparative scorecard.
No individual framework (LangChain, CrewAI, etc.) has a published standalone accuracy or savings figure — only the 7-framework aggregate from the August 12, 2026 benchmark.
Quality and cost have not yet been measured together on the same request cohort. Experiment 2 measured cost only. Separate prior evaluations produced accuracy results in the 96–100% range on different datasets. A future experiment will measure both dimensions on the same requests.
FAQ
CommonQuestions.
Which agent frameworks does CARDIAC-PURR AI Control Plane integrate with?
Seven frameworks: LangChain, LlamaIndex, AutoGen, CrewAI, Haystack, DSPy, and Semantic Kernel. Integration mechanics differ by framework — LlamaIndex, for example, is integrated via its ReAct and FunctionTool agent interfaces. An OpenAI-compatible endpoint is also available for applicable frameworks.
Does the router pick the model automatically, or can I choose?
Both. model="auto" lets the router select a tier automatically. You can also force an explicit tier with model="small", model="medium", or model="large", or override per-call with force_tier in router.query().
What does Experiment 1 (the agentic benchmark) actually measure?
A 350-request factual-QA workload (50 requests across each of the 7 frameworks), run August 12, 2026, entirely on the small tier via Claude Haiku 4.5, with caching present. It answers: does the router correctly select and route across seven agent frameworks on a factual-QA workload? It does not measure multi-tier routing, the economic cost vs a LARGE-tier reference, or cache-cold performance.
What does Experiment 2 (the economic benchmark) measure?
100 randomly selected technical queries run in September 2026 (build V120) with cache bypassed and max_tokens=2048 applied to both arms. It answers: what is the observed model inference cost difference between CARDIAC-PURR routing (model=auto) and a realized LARGE-tier reference on a technical workload? It does not measure answer quality on this cohort, routing overhead as a separate latency cost, or failure-path economics — 0 cascades and 0 retries were recorded.
Is this the same benchmark as the AI Control Plane benchmarks page?
No. The AI Control Plane benchmark measures routing accuracy and cost savings across 9 providers on general enterprise queries, run July 12, 2026. The experiments on this page are separate, framework-specific and economic studies run on different dates against different workloads.
What models were used in the benchmarks, and how do they relate to the current platform?
Experiment 1 (August 12, 2026) used claude-haiku-4-5-20251001 at the small tier. Experiment 2 (September 2026, build V120) routed to Google Gemini 2.5 Flash-Lite (93 queries), Gemini 2.5 Flash (4 queries), and Gemini 2.5 Pro (3 queries) via model=auto; the LARGE-tier reference arm ran primarily Gemini 2.5 Pro. The current production platform's large tier routes to claude-opus-4-8. Benchmark results and current platform configuration are separate — do not compare them as if they were the same test.
Evaluate Agentic Routing on Your Own Workflows.
Technical documentation and a scoped evaluation plan are provided at the start of every engagement.