Repository navigation
Known Limitations
This page documents current constraints so users can design around them.
LightAgent uses OpenAI-compatible tool schemas, but not every compatible provider implements function calling equally well.
Symptoms:
- the model ignores tools
- arguments are missing or malformed
- streaming tool-call chunks differ from OpenAI behavior
Mitigation:
- test the provider with
example/02.tools_agent.py - use
result_format="object"andtrace=True - keep tool schemas small and explicit
Default agent.run() is not a durable task. Use LightFlow checkpoints for fixed workflows, LightDAG for dynamic verified graphs, or ExecutionService for reconnectable long-running tasks. There is no generic LightAgent.resume(run_id).
Mitigation:
- make high-value tools idempotent
- persist external side effects in your application layer
- store important intermediate outputs outside the agent process
v1.0.0 durable execution is single-host and does not guarantee exactly-once external side effects.
LightSwarm is useful for simple agent handoff, but it is not a deterministic workflow engine.
Use LightFlow or LightDAG when you need explicit dependencies, verification, or durable state.
MCP integration registers tools from enabled servers and calls them through the configured transport. Long-lived enterprise MCP operation may need additional supervision in the application layer.
Mitigation:
- test MCP servers independently
- keep server config small
- disable unused servers
- add retries and health checks outside the agent when needed
Legacy arbitrary-Python tools are not registered by default and are blocked in model/tool dispatch. Direct trusted application calls are not sandboxed; see the v1.0.0 safety guide.
Mitigation:
- do not expose them to arbitrary users without isolation
- restrict filesystem and network access
- add approval for high-risk actions
- review logs for side effects
LightAgent provides a small memory protocol and MemoryPolicy, but the backend remains responsible for persistence, provenance, deletion, access control, and conflict handling.
Mitigation:
- use namespaces
- store provenance metadata
- filter by user, tenant, agent, and trust level
- treat retrieved memories as untrusted context
Default Skill discovery is not a trust approval mechanism. v1.0.0 supports explicit Provider attestation and host trust approval, including attestation for Skill bundles when the host opts in.
Mitigation:
- review third-party Skills
- inspect scripts before execution
- keep Skills in trusted repositories
create_tool() generates code from a prompt. Generated code can be incomplete, unsafe, or incompatible with production requirements.
Mitigation:
- review generated imports and side effects
- add tests
- run static checks
- avoid committing secrets
Trace model request events avoid full prompts by default, but tool arguments and outputs can contain sensitive values.
Mitigation:
- redact sensitive values before sharing traces
- disable tracing for sensitive workloads unless needed
- keep logs and trace exports access-controlled
LightAgent Wiki - see the repository, releases, and issues.
- Installation and Your First Agent
- Tools and Integrations
- Memory, Sessions, and RAG
- Workflows and Multiple Agents
- Agent Backends and Codex
- Durable Execution, Jobs, and Subagents
- Safety, Observability, and Evaluation
- Usage, Cost Estimates, and Budgets
- Deployment, Migration, and Verification
- 安装与第一个 Agent
- 工具和集成
- 记忆、会话与 RAG
- 工作流与多 Agent
- 执行后端与 Codex
- 持久执行、Jobs 与 Subagents
- 安全、观测与评测
- 用量统计、费用估算与预算
- 部署、升级与验收
- Quick Start
- Core Concepts
- API Reference
- Examples Cookbook
- Migration Guide
- Tools
- Tool Generator
- Memory
- MCP
- Skills
- LightFlow
- Tree of Thought
- Self-Learning
- Multi-Agent
- Tracing and Debugging
- Langfuse Observability
- Model Providers
- browser-use Integration
- Testing and CI
- Deployment Guide
- Architecture
- Security
- Known Limitations
- FAQ
- FAQ 中文
- Roadmap
- Release Process
- Contributing