Competitive Feature Matrix¶
How Bernstein compares to other multi-agent frameworks, as of April 2026.
Feature comparison¶
| Feature | Bernstein | CrewAI | AutoGen | LangGraph | OpenAI Agents SDK | Google ADK |
|---|---|---|---|---|---|---|
| Model agnostic | Yes — any CLI agent (Claude Code, Codex, Gemini CLI, Kiro, OpenCode, Qwen) | Yes — LiteLLM integration | Yes — supports OpenAI, Azure, local models | Yes — via LangChain model abstraction | No — OpenAI models only | No — Gemini models only |
| CLI agent support | Core design — wraps existing CLI tools directly | No — Python API agents only | No — Python API agents only | No — Python API agents only | No — API-based agents only | No — API-based agents only |
| File-based state | Yes — .sdd/ directory, no database required | No — in-memory or requires external storage | No — in-memory conversation state | Requires checkpointer backend (SQLite, Postgres) | No — in-memory | No — in-memory or cloud state |
| Self-evolving | Yes — --evolve mode analyzes metrics, proposes and executes improvements | No | No | No | No | No |
| Multi-provider routing | Yes — routes tasks to different providers based on complexity, cost, health, quota, and role-pinned provider policy | Partial — can assign different models per agent | Partial — can configure per-agent models | Yes — can assign models per node | No — single provider | No — single provider |
| Provider failover | Yes — typed rate-limit detection retries alternate providers and requeues orphaned work safely | Partial — user-defined retries | Partial — custom logic required | Partial — graph author must implement | No | No |
| Process isolation | Yes — each agent runs in its own process and git worktree | No — threads or async within one process | Partial — agents share a Python runtime | No — runs within one Python process | No — API calls within one process | No — runs within one process |
| Cost optimization | Yes — epsilon-greedy bandit learns cheapest viable model per task type, model cascade on failure | No — manual model assignment | No — manual model selection | No — manual model selection | No — single model pricing | No — single model pricing |
| Deterministic orchestrator | Yes — zero LLM tokens on coordination | No — LLM-driven agent delegation | No — LLM-driven conversation routing | Partial — graph structure is deterministic, but nodes use LLMs | No — LLM-driven handoffs | No — LLM-driven orchestration |
| Multi-turn conversation | No — agents are short-lived (1-3 tasks, then exit) | Yes — agents converse in roles | Yes — strong multi-turn, multi-party conversations | Yes — stateful conversation graphs | Yes — built for conversation handoffs | Yes — multi-turn with session state |
| Visual workflow builder | No | No | AutoGen Studio provides a UI | LangGraph Studio provides a UI | No | No |
| Ecosystem / integrations | Adapters for major CLI agents; MCP server registry | LangChain ecosystem, many tool integrations | Microsoft ecosystem, Azure integration | Full LangChain/LangSmith ecosystem | OpenAI platform (tools, retrieval, code interpreter) | Google Cloud, Vertex AI, A2A protocol |
| Open source license | Apache 2.0 | Apache 2.0 | Apache 2.0 (with CC-BY-SA docs) | MIT | MIT | Apache 2.0 |
Where Bernstein is different¶
CLI-native, not API-native. Most frameworks require you to build agents in Python using their SDK. Bernstein wraps CLI tools you already have installed. If you can run claude, codex, gemini, kiro, or opencode in your terminal, Bernstein can orchestrate it. No new abstractions to learn, no vendor SDK to import.
No LLM tokens wasted on coordination. The orchestrator is deterministic Python code. Task assignment, scheduling, health checks, and retries are all regular control flow. LLM calls happen only inside the agents doing actual work. This is a deliberate architectural choice — LLM-based schedulers are expensive, non-deterministic, and hard to debug.
Agents are disposable. Each agent spawns fresh, works in an isolated git worktree, and exits. No context window drift across tasks. No accumulated hallucinations. The janitor independently verifies each result (tests pass, files exist, no regressions) before merging. Response-cache reuse is also verification-gated, so cached completions cannot bypass that safety bar.
Cost optimization is learned, not configured. The epsilon-greedy bandit tracks which model gives the best success-rate-to-cost ratio for each task type. It starts by exploring, then converges on the cheapest model that still meets quality thresholds. Model cascade (cheap to expensive) handles failures automatically.
Self-evolution is built in. Running bernstein --evolve analyzes metrics from past runs, identifies improvement opportunities, and generates upgrade proposals. The system can improve its own prompts, routing logic, and templates over time.
Where competitors are stronger¶
Multi-turn agent conversations. AutoGen and CrewAI are purpose-built for agents that discuss problems back and forth. Bernstein's short-lived agent model is intentionally simpler but cannot replicate multi-party brainstorming or negotiation patterns.
Ecosystem breadth. LangGraph inherits the full LangChain ecosystem — hundreds of tool integrations, document loaders, vector stores. Google ADK connects natively to Google Cloud services and supports the A2A (Agent-to-Agent) protocol. If you need deep integration with a specific platform, these frameworks have a head start.
Visual development. AutoGen Studio and LangGraph Studio offer visual workflow builders for designing and debugging agent pipelines. Bernstein is CLI-first with no GUI.
Permissive licensing. All frameworks in this comparison, including Bernstein, use permissive open source licenses (Apache 2.0 or MIT).