related-work.md

Related work: efficiency + agentic coding

This is a curated reading list for the problems Happy Paths targets:

• repeated wrong-turns in agentic coding loops
• cross-session / cross-developer reuse (memory)
• correctness vs cost/time ("thinking budget") tradeoffs
• realistic evaluation of software engineering agents
• developer productivity measurement

Happy Paths is not trying to be a research project, but we do want our claims (and our measurement choices) to be legible to people who have followed this literature.

Software engineering agents + benchmarks

• SWE-bench (2023): real GitHub issues as a benchmark.

  • Paper: https://arxiv.org/abs/2310.06770
  • Website/viewer: https://www.swebench.com/

• SWE-bench Verified (2024): tighter verification harness + public results.

  • Post: https://openai.com/index/introducing-swe-bench-verified/

• SWE-agent (NeurIPS 2024): agent-computer interfaces + guardrails for SWE-bench-style tasks.

  • Paper: https://arxiv.org/abs/2405.15793

• Agentless: an “agentless” approach to solving software development problems.

  • Repo: https://github.com/OpenAutoCoder/Agentless

• SWE-SEARCH (ICLR 2025): using search/retrieval to improve SWE performance.

  • Paper (ICLR proceedings PDF): https://proceedings.iclr.cc/paper_files/paper/2025/file/a1e6783e4d739196cad3336f12d402bf-Paper-Conference.pdf

• SWE-Effi (2025): re-evaluates SWE agent effectiveness under resource constraints.

  • Paper: https://arxiv.org/html/2509.09853v2

More agentic coding benchmarks

• RepoBench (ICLR 2024): repository-level code completion.

  • Repo: https://github.com/Leolty/repobench

• AgentBench (ICLR 2024): broader benchmark for LLMs as agents (not coding-only).

  • Repo: https://github.com/THUDM/AgentBench

• ABC-Bench (2026): agentic backend coding tasks.

  • Paper: https://arxiv.org/abs/2601.11077

• ProjDevBench (2026): end-to-end project development benchmark for coding agents.

  • Paper: https://arxiv.org/html/2602.01655

• ACE-Bench (2026): agentic coding in end-to-end development of complex features.

  • OpenReview: https://openreview.net/forum?id=41xrZ3uGuI

Context engineering + memory for agents

• Structured Context Engineering for File-Native Agentic Systems (2026): schema accuracy + multi-file navigation.

  • Paper: https://arxiv.org/abs/2602.05447

• Reflexion (2023): language agents that self-improve via verbal feedback.

  • Paper: https://arxiv.org/abs/2303.11366

• MemGPT (2023): explicit external memory management for LLM agents.

  • Paper: https://arxiv.org/abs/2310.08560

• A-Mem: Agentic Memory for LLM Agents (2025).

  • Paper: https://arxiv.org/pdf/2502.12110

• Voyager (2023): lifelong skill library + curriculum in an open-ended environment.

  • Paper: https://arxiv.org/abs/2305.16291

Inference-time compute / “thinking budget” efficiency

These are relevant because “agentic coding” often fails not for lack of capability, but because the loop burns too much time/tokens getting to the right state.

• Scaling LLM Test-Time Compute Optimally… (2024).

  • Paper: https://arxiv.org/abs/2408.03314

• Token-Budget-Aware LLM Reasoning (ACL Findings 2025).

  • Paper (PDF): https://aclanthology.org/2025.findings-acl.1274.pdf

• Steering LLM Thinking with Budget Guidance (2025).

  • Paper: https://arxiv.org/html/2506.13752v1

• ThinkPrune: pruning long chain-of-thought.

  • Paper (PDF): https://openreview.net/pdf/367c67d290b7b9d4fd60db37058f764b39c667b2.pdf

Developer productivity studies (humans + assistants)

• The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (2023).

  • Paper: https://arxiv.org/abs/2302.06590

• GitHub’s own writeup of productivity/happiness study results:

  • https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-on-developer-productivity-and-happiness/

Human factors: comprehension, review, and mental models

A recurring theme in community discussion is that assistants shift the bottleneck from writing to reviewing / understanding (especially for large diffs). We want our metrics to reflect that reality.

• Towards Decoding Developer Cognition in the Age of AI Assistants (2025).

  • Paper: https://arxiv.org/html/2501.02684v1

• Understanding user mental models in AI-driven code completion tools (2025).

  • Link: https://www.sciencedirect.com/science/article/pii/S1071581925002058

• Human-AI Experience in Integrated Development Environments: A systematic literature review (2025).

  • Paper: https://arxiv.org/html/2503.06195v1

Community discussion (Hacker News + Reddit)

We also track practitioner discussion to understand what actually bottlenecks engineering teams when they adopt these tools.

Hacker News threads

• SWE-agent open source (benchmark realism, “bug report quality” debates):

  • https://news.ycombinator.com/item?id=39907468

• Devin announcement (scope limits, “90% correct is not good enough”):

  • https://news.ycombinator.com/item?id=39679787

• Aider thread (API cost, large-task iteration loops):

  • https://news.ycombinator.com/item?id=39995725

• Claude Code (multi-session orchestration, workflow discussion):

  • https://news.ycombinator.com/item?id=43307809

Reddit threads

• Claude Code velocity discussion: bottleneck shift from typing to understanding/review.

  • https://www.reddit.com/r/ClaudeAI/comments/1osv7is/using_claude_code_heavily_for_6_months_why_faster/

• AI coding assistant comparisons (Cursor/Aider/Cline/Copilot, etc.):

  • https://www.reddit.com/r/ClaudeAI/comments/1izmyps/claude_cursor_aider_cline_or_github_copilotwhich/

---

If you think a paper/thread belongs here (especially anything that quantifies cost/time or failure modes), please open a PR adding it.