Self improving agents

.python-version

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

human-seeded-evals/app/agent.py

@@ -1,35 +1,43 @@
 from __future__ import annotations as _annotations
 
+import os
 from dataclasses import dataclass
 from datetime import datetime
 
 from pydantic_ai import Agent, RunContext
 
 from .models import TimeRangeInputs, TimeRangeResponse
+from .self_improving_agent import Coach, SelfImprovingAgentModel
 
 
 @dataclass
 class TimeRangeDeps:
     now: datetime
 
 
-instrunctions = "Convert the user's request into a structured time range."
+system_prompt = "Convert the user's request into a structured time range."
 time_range_agent = Agent[TimeRangeDeps, TimeRangeResponse](
     'anthropic:claude-sonnet-4-0',
     output_type=TimeRangeResponse,  # type: ignore  # we can't yet annotate something as receiving a TypeForm
     deps_type=TimeRangeDeps,
-    instructions=instrunctions,
+    system_prompt=system_prompt,
     retries=1,
 )
 
 
-@time_range_agent.instructions
+def get_coach() -> Coach:
+    logfire_read_token = os.environ['LOGFIRE_READ_TOKEN']
+    return Coach('time_range_agent', logfire_read_token)
+
+
+@time_range_agent.tool
 def inject_current_time(ctx: RunContext[TimeRangeDeps]) -> str:
     """Add the user's current time and timezone in the format 'Friday, November 22, 2024 11:15:14 PST' to context."""
     return f"The user's current time is {ctx.deps.now:%A, %B %d, %Y %H:%M:%S %Z}."
 
 
 async def infer_time_range(inputs: TimeRangeInputs) -> TimeRangeResponse:
     """Infer a time range from a user prompt."""
-    result = await time_range_agent.run(inputs.prompt, deps=TimeRangeDeps(now=inputs.now))
+    model = SelfImprovingAgentModel('anthropic:claude-sonnet-4-0')
+    result = await time_range_agent.run(inputs.prompt, deps=TimeRangeDeps(now=inputs.now), model=model)
     return result.output

human-seeded-evals/app/main.py

@@ -1,11 +1,19 @@
+from datetime import datetime, timezone
+
 import logfire
 from fastapi import FastAPI
+from pydantic import BaseModel
 
-from .agent import infer_time_range
+from .agent import get_coach, infer_time_range
 from .models import TimeRangeInputs, TimeRangeResponse
+from .self_improving_agent import ModelContextPatch
 
 logfire.configure(environment='dev')
+
 logfire.instrument_pydantic_ai()
+logfire.instrument_httpx(capture_all=True)
+coach = get_coach()
+
 
 app = FastAPI()
 logfire.instrument_fastapi(app)
@@ -14,3 +22,36 @@
 @app.post('/api/timerange')
 async def convert_time_range(time_range_inputs: TimeRangeInputs) -> TimeRangeResponse:
     return await infer_time_range(time_range_inputs)
+
+
+class Field(BaseModel):
+    id: str
+    text: str
+
+
+@app.get('/api/context')
+def get_agent_context() -> list[Field]:
+    coach_fields = coach.get_fields() or []
+    fields = [Field(id=f.key, text=f.current_prompt or '') for f in coach_fields]
+
+    if patch := coach.get_patch():
+        for field in fields:
+            if new_text := patch.context_patch.get(field.id):
+                field.text = new_text
+
+    return fields
+
+
+class PostFields(BaseModel):
+    fields: list[Field]
+
+
+@app.post('/api/context')
+def post_agent_context(m: PostFields):
+    context_patch = {f.id: f.text for f in m.fields if f.text}
+    coach.update_patch(ModelContextPatch(context_patch=context_patch, timestamp=datetime.now(tz=timezone.utc)))
+
+
+@app.post('/api/context/update')
+async def post_update_agent_context():
+    await coach.run()