A framework for orchestrating tool-calling language model agents — routing, sandboxing, evaluating — that ships under an open license.
Scored on 5 universal + 6 category criteria. Evidenced = 3 pts, asserted = 1 pt, no claim = 0.
| Vendor | Score | Pricing transparency | Integration breadth | Data portability | Compliance posture | Documentation quality | License obligations | Tool-calling protocol | Sandbox & isolation model | Eval harness maturity | Model agnosticism | Self-host footprint |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LangChain | 58% | — | ✓ | ○ | — | ✓ | ✓ | ✓ | — | ✓ | ✓ | — |
| CrewAI | 24% | — | ○ | — | — | ○ | ✓ | ○ | — | ○ | ○ | — |
| Delta Projects | 18% | — | ✓ | — | ✓ | — | — | — | — | — | — | — |
| BidSwitch | 12% | — | ✓ | — | — | ○ | — | — | — | — | — | — |
| AdKernel | 12% | ✓ | ○ | — | — | — | — | — | — | — | — | — |
| Smadex | 9% | — | — | — | — | ✓ | — | — | — | — | — | — |
| Adelphic (Viant) | 6% | ○ | — | — | ○ | — | — | — | — | — | — | — |
| Kayzen | 3% | — | — | ○ | — | — | — | — | — | — | — | — |
Moving from linear chain-of-thought to graph-based agent orchestration unlocks parallel tool execution and conditional branching without sacrificing observability.
Role-based multi-agent crews outperform single-agent approaches by 40–60% on complex multi-step tasks requiring diverse tool access.
| Criterion | What it separates | Evidence |
|---|---|---|
| License obligations | MIT / Apache-2.0 / AGPL — what the license requires of a production deployment | doc · doc |
| Tool-calling protocol | MCP, function-calling, custom — and whether arbitrary tools plug in | doc · tested |
| Sandbox & isolation model | How agent execution is contained — containers, V8 isolates, process-level | doc · tested |
| Eval harness maturity | Built-in evaluation tooling — benchmarks, scorers, test fixtures | doc · tested |
| Model agnosticism | Number of model providers supported without a custom adapter | integration list · doc |
| Self-host footprint | What it takes to run — container image size, dependencies, minimum infra | doc · tested |