smolagents

What smolagents does

smolagents is Hugging Face's compact Python framework for agents, centered on CodeAgent. Rather than emitting one JSON tool call at a time, a model can write code that invokes tools, loops over results, performs calculations, and handles intermediate values. The repository says the agent logic fits in roughly 1,000 lines, making it easier to inspect than a large orchestration stack. ToolCallingAgent remains available for conventional structured actions.

The current repository supports Hugging Face providers, local transformers or Ollama, LiteLLM, OpenAI-compatible endpoints, Azure OpenAI, and Amazon Bedrock. Tools can come from MCP servers, LangChain, or Hugging Face Spaces. The library is free under Apache 2.0, but inference and isolated execution are not. The README names Blaxel, E2B, Modal, and Docker for sandboxing. Restrict networking, mounts, processes, time, and secrets; never run generated Python directly on a sensitive host. Evaluate 100 tasks and record completion, unsafe-action blocks, tool errors, token use, wall time, and correction effort.

Pricing, alternatives, and fit

Published pricing is: The library is free under Apache 2.0; model inference, sandbox providers, compute, storage, and operations cost extra. Pricing should be evaluated as total cost per accepted result, not only the headline fee. Include model calls, external APIs, compute, storage, retries, human review, implementation, and support. Run a capped pilot with representative peak volume and record successful completions, p50 and p95 latency, correction minutes, failures by cause, and spend per approved output.

Relevant comparisons are LangChain, CrewAI, Microsoft AutoGen, E2B. Compare them on the same inputs because category labels hide differences in hosting, permissions, integrations, observability, and exportability. The best choice is the one that meets a written quality and recovery threshold on the buyer's own workload, not the one with the longest feature list.

Pros, cons, and implementation guidance

Concrete strengths include Small abstraction layer is inspectable; Code actions compose several operations; Model- and tool-agnostic design; Apache 2.0 permits commercial modification. Important limitations include Model-written code demands sandboxing; Teams own deployment and observability; Adapters require version-aware maintenance; Code agents are harder to govern than narrow schemas. These tradeoffs need owners. Security review should cover SSO, role-based access, encryption, secrets, audit logs, subprocessors, retention, deletion, incident response, and whether customer data trains shared models. Technical review should test rate limits, timeouts, retries, idempotency, concurrency, versioning, partial failure, export, and recovery after interruption.

Practical use cases are Research plus computation agents; Sandboxed data-analysis agents; MCP tool consumers; Agent-loop education. Start with one bounded workflow whose current time, cost, and error rate are known. Use at least 50 real examples plus malformed and adversarial inputs. Keep a manual fallback until the system repeatedly meets the acceptance bar. For tools that execute actions, use least privilege and require human confirmation for money movement, publication, deletion, account changes, and external messages. For generated media or extracted documents, preserve provenance and source references so reviewers can verify consequential output.