Self-Improving Agents

Started 9/7/2025

Designing and iterating on autonomous agents that can improve themselves through tight, measurable loops. The goal is to turn agent behaviors and fixes into a continual dataset that raises capability and reliability over time.

Goals:

• Close the loop: generate → evaluate → critique → patch → verify → deploy.
• Capture high-signal traces, failures, and fixes to fuel fine-tuning/RM updates.
• Keep everything measurable: clear success criteria, regression tests, and dashboards.

Components:

• Orchestrator: task router + run manager with structured traces and artifacts.
• Evaluators: unit/integration tests, adversarial prompts, safety/guardrail checks.
• Memory/Data: append-only store of runs, critiques, diffs, and outcomes; retrieval for planning.
• Skills/Tools: code, shell, web, structured APIs; composable tool-using plans.
• Training: small fine-tunes (SFT/DPO) and reward modeling from curated traces.
• Sandbox: deterministic runners with fixtures for reproducible evaluation.

Planned:

• Seed task suites (coding, data wrangling, web research) with clear metrics.
• Failure harvesting: auto-minimize and save failing traces as regression tests.
• Self-edit flow: draft change proposals (PR-style diff), run eval battery, gate on pass.
• RM-assisted critique to prioritize high-impact fixes and reduce regressions.
• Dashboard: success rates, time-to-fix, and distribution shift alerts.

Progress:

2025-09-07 — Initial scope and plan

• Added project entry and outlined components and milestones.
• Next: scaffold evaluator harness and a minimal seed task set.

Links (todo): Repo + design doc.