deepagents
Facts
- Openness
- Open source
- Licence
- MIT
- Language
- Python
- Pinned version
4394bcd(2026-09-30)- Topics checked
- 31 of 33
- Kind of agent
- General framework
What it does notably well
- Message model
Before-agent pass inserts error ToolMessages for unanswered AI tool-call ids.
- Human approval
When the caller supplies interrupt-mode filesystem permission rules, graph assembly turns them into conditional LangChain HITL interrupts: per-tool entries whose path predicates…
- MCP and auth
Allocate stable, provider-safe MCP tool names from server and original tool identity, adding a hash for sanitized or long names and an order-independent collision suffix.
- Prompt caching
When SDK memory sources are configured with an Anthropic model, prompt-caching middleware marks the preceding system prefix before MemoryMiddleware appends memory and marks its…
- Self-improvement
Memory is treated as untrusted data in the prompt itself: the memory guidelines say agent_memory is file data that may be outdated, incorrect or written by someone other than the…
- Skills
Metadata is cached in graph state per thread and can be explicitly reloaded by setting skills_metadata to None.
See how deepagents does on each topic
The topic-by-topic verdicts, the code lines and a side-by-side comparison are in the app. Sign in with Google; new accounts are approved by hand.
Independent analysis, not affiliated with deepagents.