Building Self-Improving Engineering Systems

Resources

• Demystifying Evals for AI Agents — Anthropic explains tasks, trials, transcripts, outcomes, grader types, and capability versus regression evaluations. Useful for constructing a proof ladder from deterministic validators through calibrated judgment, rather than promoting a system because its own explanation sounds successful. https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents

• OpenAI Trace Grading — Documents evaluating the decisions and tool interactions inside an agent run, not only its final response. Use it to turn recurring routing errors, bad tool choices, and missed escalation opportunities into targeted improvement candidates. https://developers.openai.com/api/docs/guides/trace-grading

• LangSmith Dataset Management — Covers evaluation examples, dataset versions, and organizing test data. Useful for turning production failures, review findings, and human corrections into reproducible regression cases tied to a known baseline. https://docs.langchain.com/langsmith/manage-datasets

• DSPy — The original framework separates language-model programs from optimizers that tune them against explicit metrics and examples. Provides a practical model for evaluating profile and instruction changes systematically instead of accumulating untested prompt edits. https://github.com/stanfordnlp/dspy

• GEPA: Reflective Prompt Evolution — The original paper explores using execution feedback and reflection to evolve prompts. Useful for a propose–test–compare loop over agent configurations, with held-out evaluations and cost tracking needed before adopting an apparent improvement. https://arxiv.org/abs/2507.19457

• Voyager — The original project demonstrates an embodied agent that builds a reusable library of executable skills through interaction and feedback. A concrete example of successful-run distillation, with the important boundary that results in Minecraft do not establish safe autonomous improvement in production engineering. https://voyager.minedojo.org/

• Reflexion — This research studies agents that store verbal reflections from feedback to improve subsequent attempts without updating model weights. Useful for distinguishing episodic correction and memory from training, and for asking whether a reflection actually helps on fresh cases. https://arxiv.org/abs/2303.11366

• Agentic Context Engineering — The original paper studies evolving context playbooks through generation, reflection, and curation. Connects observed failures and successful strategies to reusable rules and context while motivating checks against stale, contradictory, or overfitted guidance. https://arxiv.org/abs/2510.04618

• Google SRE’s Postmortem Culture — Explains blameless incident analysis and actionable follow-up rather than collecting failure reports for their own sake. Use it to structure human decisions and failure evidence before changing agent tools, environments, routing policies, or permissions. https://sre.google/sre-book/postmortem-culture/

• Canarying Releases — Google’s SRE workbook explains comparing a limited rollout against a control and deciding whether to continue or roll back. Adapt the mechanism to new models, tools, workflows, and agent profiles, measuring quality alongside latency, cost, and operational failures. https://sre.google/workbook/canarying-releases/

• GitHub Protected Branches — Documents required reviews, status checks, and restrictions around merging changes. Use it to keep proposed skill, rule, workflow, and agent updates behind independent evaluation and human approval rather than letting a system authorize its own promotion. https://docs.github.com/en/repositories/configuring-branches-and-merges-in-your-repository/managing-protected-branches/about-protected-branches

• Git Revert — The official reference describes recording a new commit that reverses an earlier change. Provides an auditable rollback mechanism for versioned control-plane improvements, while distinguishing configuration rollback from repairing side effects already produced by the changed system. https://git-scm.com/docs/git-revert