The 0.2.0 harness update and subsequent Lovable compatibility work were researched on 2026-07-17. These sources inform the repository design; they are not loaded by the runtime skill.
- Build skills: progressive disclosure, focused skill scope, explicit inputs/outputs, and trigger testing.
- Codex best practices: durable repository context, verification, review, and reusable workflows.
- Subagents: context isolation, summary-based handoffs, and caution around parallel writes.
- Using PLANS.md for multi-hour problem solving: self-contained living plans, observable outcomes, progress, decisions, and validation evidence.
- Trace grading: reproducible evaluation of agent decisions, tool calls, and outcomes.
- Build an Agent Improvement Loop with Traces, Evals, and Codex: traces-to-feedback-to-evals improvement flywheel.
- Agent internet access: prompt injection, exfiltration, malware, license risk, and network minimization.
- Harness engineering: concise repository maps, executable constraints, and harness maintenance. This is an internal production case study, not a controlled benchmark.
- Demystifying evals for AI agents, 2026-01-09: realistic tasks, multiple trials, balanced cases, outcome graders, regression suites, and transcript review.
- Effective context engineering for AI agents, 2025-09-29: finite attention budgets, minimum sufficient context, just-in-time retrieval, compaction, notes, and subagents.
- Effective harnesses for long-running agents, 2025-11-26: feature ledgers, incremental work, clean session boundaries, baseline checks, and end-to-end verification.
- Writing tools for agents, 2025-09-11: distinct high-impact tools, semantic interfaces, token-efficient responses, and eval-driven tool improvement.
- How we built our multi-agent research system, 2025-06-13: delegation contracts, effort scaling, artifact handoffs, observability, and the cost/coordination limits of multi-agent work.
- Building effective agents, 2024-12-19: start simple, match architecture to task predictability, use environmental feedback, and add complexity only when justified.
- Define reusable instructions with skills: workspace skill imports, the Agent Skills
SKILL.mdconvention, ZIP structure, bundled-file limits, automatic activation, and slash-command invocation.
- Andrej Karpathy, autoresearch, 2026-03: fixed evaluator, bounded trials, one mutable surface, one metric, and keep/discard iteration. This is a reference implementation, not evidence that the pattern generalizes to durable product code.
- Mitchell Hashimoto, My AI Adoption Journey, 2026-02-05: turn recurring agent mistakes into repository guidance or deterministic tools. Practitioner experience.
- Armin Ronacher, The Coming Loop, 2026-06-23: autonomous loops work well for experiments, ports, research, and mechanically verifiable tasks, with caution for long-lived code. Practitioner judgment.