Tim's Wiki

Coding Agents and AI Tools

Properties2
descriptionLandscape of AI-powered coding agents, orchestration frameworks, harnesses, and supporting tools for software development.
tagsai, tools, software-engineering, agents

7 min read

Coding Agents and AI Tools

Overview

The coding agent ecosystem has evolved rapidly from simple code completion (GitHub Copilot) to autonomous software engineering platforms. The space now spans IDE-integrated agents, CLI-based agents, orchestration layers, harness frameworks, and supporting infrastructure like MCP servers and spec-driven development tools.

Coding Agents

IDE-Integrated

  • Cursor — AI-first code editor with model pricing and CLI
  • Windsurf — AI coding environment
  • Void (paused) — open source Cursor alternative (YC-backed)
  • AmpCode — multi-model, CLI only

CLI Agent Harness

Design-Focused

  • Stitch — design-focused agent from Google

Orchestration Frameworks

Tools for coordinating multiple agents or models:

  • Conductor — agent orchestration platform
  • Agent Swarm — framework for AI coding agents
  • Agent Teams (Claude Code) — multi-agent collaboration within Claude
  • AI Stack — AI agent stack
  • Chorus — agent harness for AI-human collaboration, inspired by AI-DLC
  • Fanbox — vibe coding cockpit with file browser, terminal, and change tracking
  • Multica — multi-agent coordination (complex setup)
  • acpx — runs heterogeneous agents through the Agent Client Protocol in a terminal, non-interactively; useful when the point is to drive several different agents uniformly rather than to supervise one

Harnesses

Lightweight runtimes that wrap coding agents with structure:

  • Powerball Harness
  • Forge

Skills Management

Systems for giving agents reusable domain knowledge:

  • PRPM — package manager for agent skills
  • OpenSkills — open skill repository
  • Skills (Vercel) — Vercel’s approach to agent skills

Spec-Driven Development

Tools that generate implementations from specifications:

  • OpenSpec
  • Spec-Kit

Open-Weight Code Models

  • StarCoder — LM trained on source code, from Hugging Face + ServiceNow
  • Stable Code — Stability AI’s code LLM
  • Code Llama — Meta’s code generation model
  • FauxPilot → OpenHands — open-source Copilot server using moyix models

Autonomous Agent Platforms

Supporting Tools

Token Efficiency

  • VIBEguard — guardrails for agent output
  • Headroom — token saving
  • RTK — token reduction toolkit

Web Fetching / Content Extraction

  • Firecrawl — web scraping for AI, good for walled sites (Reddit, Twitter)
  • Exa — AI search engine
  • Jina — content conversion for AI
  • ScrapeGraph AI — AI-powered scraping
  • Kadoa — web fetching layer for finance

MCP Servers

Model Discovery

Earlier Generation (2022–2023)

Kept for lineage — most are superseded by the CLI agents above, but they show what the first wave optimized for (search and explanation, not autonomous editing):

  • OpenDevin → OpenHands — platform for autonomous software engineers (paper)
  • Bloop AI — semantic code search over a codebase with GPT-4
  • sketch — code assistant that reads data content, not just code
  • Adrenaline — Codex-powered debugging assistant
  • stackoverflow.gg — ChatGPT answers for every Stack Overflow question, including unanswered ones

Chat as pair programmer (first-hand, 2023)

A contemporaneous account of using bare ChatGPT — no IDE plugin, no file access — as a coding partner. Worth keeping because the listed failures are exactly what the harness generation was built to fix, which dates the problems rather than dismissing them.

The workflow was crude by necessity: describe anything visual in text, because it could not read images; paste raw source into the input box, unformatted, since there was no rich editor. “If that were a colleague, they wouldn’t be happy — ChatGPT doesn’t complain.”

What worked: whole React components, project structure suggestions (stateful and stateless components in separate folders), UI guidance grounded in real design principles, and — rated highest — code review, catching many general issues in one pass and judged better than the GitHub review-quality integrations of the time.

What failed, and where it went:

2023 failureWhat fixed it
forgot its own earlier decisions mid-conversationcontext management and longer windows
suggested outdated package versions (Chakra UI 0.8.0)tool use, real file and network access
type errors and missing imports ESLint would catchagents that run the linter and iterate on the result

The last row is the durable point: the gap was never reasoning about code, it was the feedback loop. An assistant that cannot run anything cannot check anything — which is why AI-Native SDLC makes “always give the agent a way to verify its own work” the central instruction rather than an optimisation.

Practitioner Insights (Boris Cherny, Claude Code)

From the Pragmatic Engineer interview with Claude Code’s creator:

  • Plan-mode, then one-shot — iterate on a plan first; once the plan is good, the model “one-shots the implementation almost every time.” Boris runs ~5 parallel Claude instances (separate checkouts) to ship 20–30 PRs/day.
  • Agentic search = glob + grep — plain model-driven glob/grep beat local vector DBs and recursive indexing (which suffered stale indexes and permission complexity). Simpler retrieval won.
  • Finish your migrations — partially-migrated, multi-framework codebases confuse both humans and models; code quality has a measurable double-digit-percent productivity impact.
  • Automate review into lint rules — log repeated review comments; once a pattern recurs 3–4 times, write a lint rule to eliminate it.
  • Prototypes replaced PRDs — build dozens of working prototypes instead of static specs/mocks.
  • Fix infra before building product — you can’t build great products on a broken foundation.

Adoption & Impact (Microsoft study, early 2026)

A Microsoft study of tens of thousands of engineers over 4 months (rolling out Claude Code and GitHub Copilot CLI) answered three questions — who adopts, who retains, and is the productivity worth the token cost:

  • Adoption spreads socially, not top-down. The strongest driver of first use was peer influence — teammates, code reviewers, and surrounding engineers already using it. Diffusion resembles a social network, not a traditional IT rollout. Implication: cultivate visible champions/opinion leaders.
  • Retention tracks coding activity, not demographics. Continued use correlates with high coding workload and active dev tasks — weakly with age, gender, level, or demographics. “Do they write code often?” predicts retention, not “who they are.”
  • ~+24% merged PRs. Adopters merged ~24% more pull requests (vs a modeled counterfactual), and the effect persisted across the full 4 months — not a novelty spike.
  • Dose-response. Higher usage frequency → larger PR gains; heavy users (≥5 days/week) improved most.
  • Limitations. Merged PR ≠ software value (quality/customer/business value may not rise in step) — the same caution as Measurement Dysfunction; results are Microsoft-specific (mature AI-coding culture) and observational (causal inference, not an RCT).

Takeaway for orgs: track adoption and retention (not license count), seed via technical champions, target high-coding-volume developers first, and build an ROI system spanning token cost, usage, output, code quality, and business value — mirroring Developer Productivity Measurement.

  • Model Context Protocol (MCP)
  • Context Engineering — designing what the harness puts in front of the model; why 80% of Claude Code’s system prompt was deleted
  • Reasoning Effort Control — the low/medium/high dial these agents run on, and what it actually costs
  • Loop Engineering — the four loop types (turn-based, goal, time, proactive) and what you hand off in each
  • AI Model Routing
  • Cloudflare OS — sandboxed per-user AI micro-apps behind capability proxies; the Code-Mode agent that writes TypeScript instead of tool calls
  • Agentic Cyber Attack — the same harness and swarm patterns, used offensively
  • Developer Productivity Measurement — how to measure the productivity impact this study estimates
  • AI-Native SDLC — the six-stage lifecycle these agents are wired into, and the artifact chain that connects the stages
  • Agent Skill Evolution — agents compiling their own execution traces into skills, and the persistent wiki that makes it compound
  • AI Inference Hardware — the silicon and the tokens-per-megawatt economics underneath all of it
  • Building Claude Code with Boris Cherny (Pragmatic Engineer)
  • Micro-Agent: Beat Frontier Models with Collaboration inside Model API

Source References

  • coding agents
  • 微软 CLI AI 编程代理研究

Graph View

Backlinks