loc bengaluru, ist | local --:-- srijanshukla18@gmail.com →
[post]/ai/may-ai-builder-newsletter

AI builder notes - May 2026

/ 5 min read· ai

AI-assisted notes from my liked-tweets feed, organized around agent workflows, browser traces, model loops, and guardrails.

Notes from my May liked tweets. Lot of people converging on the same few patterns: explicit workflows over open-ended agents, pairing expensive planners with cheap executors, and replacing bloated MCP servers with plain CLIs/SDKs.

Main takeaways from May

  1. Define the workflow before picking the agent. “Review this repo” is a useless prompt. “Check these auth routes, cite file + line number, run the auth test suite, stop after the first working exploit” is something an agent can actually execute. Claude Code’s dynamic workflows feature is built around this - you describe the multi-step job and it writes a JS orchestrator script to run the parallel and sequential steps. [1]
  2. Expensive model plans, cheap model executes. Use your best frontier model (like GPT-5.5 xhigh or Opus) to investigate the repo and write the implementation plan, then hand the isolated tasks to cheaper/faster models (Composer 2.5, GLM 5.1, etc.) once the tests and specs are clear. [2] [3]
  3. MCP tool schemas eat way too many tokens. Wild benchmark from monday.com’s GraphQL API: the exact same task took 15k tokens using SDK/Code Mode vs 158k tokens through an MCP server. Default to CLIs or small scripts locally; only use MCP when you genuinely need a remote wrapper. [4] [5]
  4. For browser agents, record the trace and shrink it into a skill. Watch the browser agent fumble through the UI once, strip out all the wasted clicks, replace slow DOM clicking with direct JS/reads where possible, and save the fast path as a SKILL.md. [6]
  5. Put security in the harness, not the system prompt. Stop conditions, permission gates, secret proxies, npm package-age delays - enforce them in code outside the model.

Workflows & long agent runs

  • Claude Code dynamic workflows: writes a JS orchestration script on the fly to fan out subagents on bounded tasks (e.g. checking 5 microservices or testing 3 candidate fixes) and merge the results. [1]
  • 30-minute Cursor review skills: giving a code-review skill a 30-minute budget so it actually follows call graphs across files instead of leaving shallow lint comments. [7]
  • “100 tool calls before answering”: prompt trick for Codex to stop it from guessing after reading 2 files - force it to keep grepping and verifying before it answers. [8]
  • Model-specific failure modes: Claude Code and Codex get stuck in different ways, so your harness retry/stop rules need to account for which model is driving. [9]

Running cheaper models in tight loops

  • Letta Code + GLM 5.1 review bot: works well coz the harness feeds it exact repo context, examples of good review comments, and a strict output schema. [3]
  • Ramp’s RL-trained spreadsheet model: beat Opus on their internal spreadsheet retrieval benchmark after narrow RL training on that exact task. [10]
  • Auto-repairing broken tool calls: Command Code shared stats on catching and fixing tens of thousands of malformed JSON / bad tool args automatically in the harness so the agent run doesn’t crash. [11]

Speeding up browser agents

Coolest browser example this month was Hermes Agent / Autobrowse on a Hacker News task:

  • First run: 102 seconds, 23 turns, $1.46
  • After distilling the trace into a reusable skill: 35 seconds, 8 turns, $0.28 [12] [6]

Other browser-agent stuff to keep an eye on: OpenAI’s Codex Chrome extension [13], BrowserCode [14], browser-harness [15], and Pi browser extensions [16].

  • Birdclaw [17] - lets agents search your Twitter archive.
  • GBrain [18] - personal recall layer for OpenClaw / Hermes setups.
  • PageIndex [19] - good reminder that plain BM25 + tree structure often beats vector embeddings for docs.
  • Jason Liu’s thread on why RAG is gonna make a comeback in ~8 months once everyone realizes stuffing 1M tokens of unindexed markdown into context doesn’t scale. [20]

Security & guardrails

Cloudflare tested Claude Mythos across 50 repos [21], and Mozilla used Mythos Preview to fix more Firefox security bugs in April than in the previous 15 months combined. [22]

One super practical guardrail you should turn on right now if coding agents run npm install or pnpm add on your machine: minimum package release age. Stops agents from pulling a freshly hijacked npm package 5 minutes after an attacker publishes it. [23] [24]

pnpm config set minimumReleaseAge 2880
npm config set min-release-age=2d

Also saw Clawvisor [25] for proxying credentials so the agent never sees raw API keys in its environment.

Other tools from the feed

  • Harness engineering guide - solid breakdown of evals, stop rules, retries, and verification loops
  • LiteParse v2 - fast Rust PDF parser for RAG/agents
  • Patter - unified wrapper for voice AI providers
  • Minions - dashboard UI for managing multiple local Hermes agents
  • OpenRouter Pareto Router - routes requests to the cheapest model that clears a quality threshold
  • OpenRouter Response Caching - saves money on repeated prefix/retry loops
  • Flue - TypeScript framework for running agents inside sandboxes with a secret proxy
  • Zero - experimental language designed for agents with JSON compiler diagnostics and explicit capability grants