If yesterday’s agents were students that crammed harder prompts and re-ran until they got lucky, the new wave is closer to apprentices who learn a craft, bottle the technique, and reuse it. A Princeton + Tsinghua team just showed a clean, pragmatic way to do that: start with a general agent, let it work real tasks, and every time it succeeds, distill the method into a reusable tool. Over time, the agent teaches itself into a domain expert.
They call the recipe Alita-G. The trick is simple enough to feel obvious in hindsight:
- When a task goes well, the agent saves not just the final answer but the procedure as an MCP tool.
- It then abstracts that tool: add parameters, strip task-specific bits, standardize the interface, write docs.
- The result goes into an MCP Box—a curated toolbox for the domain.
- At question time, a retriever matches the incoming task to tool descriptions/use cases and picks what to call.
- A minimal runtime stack—Task Analyzer → MCP Retriever → MCP Executor—does the rest.
Why this matters: prior “self-evolving” systems mostly tweaked prompts, tuned a module, or retried failures. Gains were narrow and brittle. Here, the unit of learning is a capability you can call again. That compounds.
The payoff shows up on benchmarks and the bill. On GAIA, the specialist built by Alita-G hits 83.03% pass@1 while using ≈15% fewer tokens than a strong baseline agent—higher accuracy with less compute. Similar patterns show up on PathVQA and HLE: the specialists beat the generalist and stay lean.
A few design notes that stood out:
- Self-harvest, then generalize. Capturing a successful trajectory as an MCP is good; abstracting it into a parameterized primitive is what makes it portable.
- RAG at the tool level. Retrieval isn’t about documents—it’s about which function to arm the agent with. Descriptions + use cases make effective retrieval keys.
- Diminishing returns are real. Running multiple harvest rounds enriches the MCP Box, but results level off around 3 rounds as near-duplicates creep in.
- Small stack, big leverage. The architecture stays compact; the “knowledge” lives in the toolbox.
Zooming out, this reframes “agent improvement” from polishing prompts to compounding a library of working skills. It’s closer to how teams build internal playbooks: keep what worked, generalize it, document it, and make it discoverable. The difference is that here the agent writes—and later retrieves—its own playbook.
If you want the technical deep dive (benchmarks, ablations, and the exact MCP/RAG mechanics), see the paper: Alita-G: Self-Evolving Generative Agent for Agent Generation.