Dumping the best stuff from my April liked tweets. Most of the feed this month was about everything around the model: harnesses, managed sandboxes, memory compaction, SKILL.md workflows, and where companies are actually making money with AI.
6 takeaways from April
- The harness matters as much as the model. Two tools calling the exact same model feel completely different coz of how they handle context packing, tool definitions, caching, compaction, permissions, and verification loops. See OpenAI Agents SDK [1], Stanford’s Meta-Harness [2],
browser-harness[3], Agent CI [4], and Passmark [5]. - Write the evals and specs first. Code generated by an agent is disposable; you can regenerate it anytime. Your test suite, assertions, and trace scorers are the actual asset. [1] [2] [5]
- Managed agent runtimes are getting good. Anthropic’s Claude Managed Agents split the setup cleanly into brain / hands / durable session log, with credentials kept outside the disposable container, OpenTelemetry built in, and $0.08 per session-hour. [6]
- Background agents need a real workspace and state. Whether it’s OpenAI Codex [7], Claude Code routines [8], or Gemini CLI subagents [9], everyone is adding cron schedules, persistent devboxes, and ways for the agent to pause and ask permission.
- Dumping markdown in a folder is not “memory”. A big folder of files is just storage. Real memory is compaction + deciding what gets injected into the context window when, without polluting the prompt. [10] [11] [12] [13]
- Boring enterprise plumbing is where the revenue is. Migrations, data cleanup, compliance checks, annotation pipelines, legacy internal tools. Wildest stat from April: HireCade hit $22M ARR at 95% margins with 5 people in 14 months, bootstrapped, mostly doing annotation services with a 30% fee. [14]
Harnesses & testing loops
- OpenAI Agents SDK [1] added built-in primitives for handoffs, trace logging, sandboxes, and long-running sessions.
- Meta-Harness from Stanford [2] treats the harness code itself as something you optimize in a loop against eval scores and execution traces.
browser-harness[3] wraps Chrome CDP so browser scripts self-heal when selectors change.- Agent CI [4] runs GitHub Actions locally so coding agents can test their own CI changes before pushing.
- Passmark [5] sits on top of Playwright for natural-language regression checks with caching and telemetry.
Managed agents & enterprise tools
- Claude Managed Agents [6]: disposable Linux containers, durable event log, credentials isolated outside the sandbox, 60% faster median TTFB.
- OpenAI Codex [7]: turning into a full dev environment with browser use, SSH/devboxes, cron automations, and screen-context memory.
- Shopify AI Toolkit [15] and Cloudflare Agent Lee [16]: giving agents scoped write access to real prod stuff (Workers, R2, DNS, store inventory, orders) behind strict permission boundaries.
Memory & SKILL.md workflows
A lot of people are building personal/team knowledge bases on top of plain markdown + git right now: GBrain [10], LLMwiki [11], Rowboat [12], Obsidian [17], and Hermes [13]. The tricky part isn’t storing the markdown, it’s getting the agent to actually search it when it doesn’t realize it’s missing context.
Also, SKILL.md is pretty much the standard way to package workflows now. Do a task once with an agent, look at the trace, delete all the dead ends, and save the working commands + rules into a SKILL.md file for next time.
A few workflow patterns from the feed worth copying:
- DB query optimization loop: seed realistic test data -> run
EXPLAIN ANALYZE-> let the agent try 10 index/query variants -> keep the fastest one. - Spec-first coding: human reviews
SPEC.mdand interface types, agents write the implementation behind the interface. - Keep an
implementation-notes.htmlor.mdlog: have the agent log architectural tradeoffs and open questions while it works so you can review the decisions quickly. - Cron + 1 LLM call > autonomous agent: if a workflow is 90% deterministic, write a normal cron script and make a single LLM call for the fuzzy part instead of burning tokens on an open-ended agent loop.
Tools worth checking out
- Hermes Agent - open-source personal agent stack with skills, memory, WhatsApp/Telegram/phone hooks, and cost tracking
- GBrain - Garry Tan’s knowledge-graph recall layer for OpenClaw / Hermes
- Cabinet - local-first open-source workspace with scheduled agents and knowledge base
- browser-harness - self-healing CDP browser wrapper
- Pi - super clean extensible coding agent harness
- HyperFrames - renders HTML animations to MP4 for agents
- Firecrawl web-agent & Firecrawl Parse - web/PDF scraping into clean markdown
- Agent CI - run GitHub Actions locally in an agent loop
- awesome-design-md -
DESIGN.mdstyle guides so agents don’t generate generic UI - Syncthing - old-school P2P file sync, great for syncing markdown/skills folders across machines without cloud lock-in