working self-improving agents

Author: samuelcolvin

.gitignore

@@ -11,3 +11,4 @@ wheels/
 .venv
 *.svg
 scratch/
+.self-improving-agent/

.python-version

@@ -1 +1 @@
-3.12
+3.13

human-seeded-evals/app/agent.py

@@ -6,13 +6,12 @@
 from datetime import datetime
 from typing import AsyncIterator
 
-from cloudkv import AsyncCloudKV
 from pydantic_ai import Agent, RunContext
 from pydantic_ai.models import Model
 
 from .models import TimeRangeInputs, TimeRangeResponse
 from .self_improving_agent import SelfImprovingAgentModel
-from .self_improving_agent_storage import CloudKVStorage
+from .self_improving_agent_storage import LocalStorage
 
 
 @dataclass
@@ -32,13 +31,14 @@ class TimeRangeDeps:
 
 @asynccontextmanager
 async def self_improving_model() -> AsyncIterator[SelfImprovingAgentModel]:
- cloudkv_read_token, cloudkv_write_token = os.environ['CLOUDKV_TOKEN'].split('.')
  logfire_read_token = os.environ['LOGFIRE_READ_TOKEN']
- async with AsyncCloudKV(cloudkv_read_token, cloudkv_write_token) as cloudkv:
- storage = CloudKVStorage(cloudkv)
- m = SelfImprovingAgentModel('anthropic:claude-sonnet-4-0', storage, logfire_read_token, 'time_range_agent')
- yield m
- await m.wait_for_coach()
+ # cloudkv_read_token, cloudkv_write_token = os.environ['CLOUDKV_TOKEN'].split('.')
+ # async with AsyncCloudKV(cloudkv_read_token, cloudkv_write_token) as cloudkv:
+ # storage = CloudKVStorage(cloudkv)
+ storage = LocalStorage()
+ m = SelfImprovingAgentModel('anthropic:claude-sonnet-4-0', storage, logfire_read_token, 'time_range_agent')
+ yield m
+ await m.wait_for_coach()
 
 
 @time_range_agent.instructions

human-seeded-evals/app/main.py

@@ -11,6 +11,7 @@
 logfire.configure(environment='dev')
 
 logfire.instrument_pydantic_ai()
+logfire.instrument_httpx()
 
 
 @asynccontextmanager

human-seeded-evals/app/self_improving_agent.py

@@ -429,9 +429,9 @@ def get_instrunctions(messages: list[ModelMessage]) -> str | None:
 
 def get_tools_fields(tools: list[ToolDefinition], prefix: str, description: str) -> Iterable[FieldDetails]:
  for t in tools:
- prefix = f'{prefix}.{escape_key(t.name)}'
- yield FieldDetails(f'{prefix}.description', description, t.description)
- yield from json_schema_fields(t.parameters_json_schema, f'{prefix}.parameters')
+ tool_prefix = f'{prefix}.{escape_key(t.name)}'
+ yield FieldDetails(f'{tool_prefix}.description', description, t.description)
+ yield from json_schema_fields(t.parameters_json_schema, f'{tool_prefix}.parameters')
 
 
 JsonSchema = dict[str, Any]

human-seeded-evals/app/self_improving_agent_storage.py

@@ -1,7 +1,10 @@
+import asyncio
 from contextlib import asynccontextmanager
 from dataclasses import dataclass
 from datetime import timedelta
-from typing import AsyncIterator
+from functools import partial
+from pathlib import Path
+from typing import AsyncIterator, Callable, ParamSpec, TypeVar
 
 from cloudkv import AsyncCloudKV
 
@@ -30,3 +33,43 @@ async def lock(self, agent_name: str) -> AsyncIterator[bool]:
  await self.cloud_kv.delete(key)
  else:
  yield False
+
+
+@dataclass
+class LocalStorage(SelfImprovingAgentStorage):
+ directory: Path = Path('.self-improving-agent')
+
+ def __post_init__(self):
+ self.directory.mkdir(exist_ok=True)
+
+ async def get_patch(self, agent_name: str) -> ModelContextPatch | None:
+ file = self.directory / f'{agent_name}.json'
+ if file.exists():
+ content = await asyncify(file.read_bytes)
+ return ModelContextPatch.model_validate_json(content)
+
+ async def set_patch(self, agent_name: str, patch: ModelContextPatch, expires: timedelta) -> None:
+ # note we're ignoring expiry here
+ file = self.directory / f'{agent_name}.json'
+ content = patch.model_dump_json(indent=2)
+ await asyncify(file.write_text, content)
+
+ @asynccontextmanager
+ async def lock(self, agent_name: str) -> AsyncIterator[bool]:
+ file = self.directory / f'lock:{agent_name}'
+ if not await asyncify(file.exists):
+ await asyncify(file.touch)
+ try:
+ yield True
+ finally:
+ await asyncify(file.unlink)
+ else:
+ yield False
+
+
+P = ParamSpec('P')
+R = TypeVar('R')
+
+
+async def asyncify(func: Callable[P, R], *args: P.args, **kwargs: P.kwargs) -> R:
+ return await asyncio.get_event_loop().run_in_executor(None, partial(func, *args, **kwargs))

human-seeded-evals/update_sia.py

@@ -0,0 +1,20 @@
+import asyncio
+
+import logfire
+from app.agent import infer_time_range, self_improving_model
+from app.models import TimeRangeInputs
+
+logfire.configure(environment='evals')
+
+logfire.instrument_pydantic_ai()
+
+
+async def main():
+ async with self_improving_model() as model:
+ with model.blocking_context():
+ with logfire.span('running infer_time_range with blocking coach'):
+ await infer_time_range(TimeRangeInputs(prompt='yesterday'), model=model)
+
+
+if __name__ == '__main__':
+ asyncio.run(main())

pyproject.toml

@@ -18,10 +18,6 @@ dependencies = [
 [dependency-groups]
 dev = ["pyright>=1.1.402", "ruff>=0.12.1", "watchfiles>=1.1.0"]
 
-
-[tool.uv.workspace]
-members = ["human-seeded-evals"]
-
 [tool.ruff]
 line-length = 120
 target-version = "py39"