AI builder notes - June 8, 2026

Went through 248 liked tweets from June 1 to June 7, 2026. Dumping the useful stuff from the week below.

Top takeaways from the week

1. Put validation inside the agent loop, not on the human. Run typechecks, lint, tests, builds, and headless browser checks before the agent hands anything back to you. Let the agent fix its own compiler/test errors first so you only review whether it built the right thing. [1] [2]
2. Throwaway verification scripts. Claude Code’s dynamic workflows can write a one-off JS harness for a specific task - like extracting every technical claim from a draft blog post, checking them against the repo files, and flagging anything that doesn’t match. [3] [4]
3. Cloud agents = distributed systems problems. Moving an agent off your laptop and onto remote VMs means dealing with pod lifecycles, stream rewinds on reconnect, state isolation, and partial output during retries. [5] [6]
4. Skills need supply-chain scanning. If agents are pulling SKILL.md packages from registries and repos, you gotta scan them for prompt injection and sketchy shell commands just like npm packages. [7] [8]
5. Give agents raw evidence, not vague descriptions. Paste the failing curl output, the stack trace, the log lines, the DB row. [9]
6. Agent memory as shared work state. Less about a personal archive of old notes, more about tracking what is currently true and what already failed so a second agent can pick up safely. [10] [11]

Backpressure in the loop

If you don’t have automated checks in the loop, you end up playing human CI: agent writes code -> you run it -> import is broken -> you paste error back -> agent fixes import -> now test fails -> repeat.

Give the harness the typechecker, linter, test runner, and browser logs directly so the loop doesn’t stop until those pass. [1]

Also liked this framing from @code_rams: treat a workflow like a team with separate steps for planning, parallel execution, and adversarial review. [12] Though tbh if your check is just running 3 bash commands, don’t overcomplicate it with a dynamic workflow - just run the 3 commands.

Cursor’s cloud agent writeup

Peter Pang’s post on moving Cursor’s desktop agent to the cloud is really good. [5]

Once the loop isn’t running on your local machine, you hit all the classic infra headaches: who owns the VM state? What happens when a pod restarts mid-run? How do retries interact with token streaming so the user doesn’t see corrupted/duplicate diffs in the UI?

They ended up using Temporal to decouple the orchestration loop from the sandbox VM lifecycle.

Skills as packages (and security headaches)

Hiten Shah had a good tweet on turning how your best engineers work into reusable skills. [13] And Vercel launched their skills.sh API with 600k+ searchable skills and project-scoped OIDC auth. [7] [14]

Obvious catch: if agents install and follow SKILL.md files automatically, someone is gonna hide malicious instructions in them. NVIDIA open-sourced SkillSpector to scan skills for hidden prompt injections, data exfiltration paths, and shadow triggers. [8] [15]

Evidence > guessing

When coding agents fail on bugs, it’s usually coz they read the source code, make up a theory in their head, and write a fix for their imaginary theory. Skip that by handing them the exact failing trace, payload, or command output upfront. [9]

Best example of this from the week: PostHog ran an overnight Autoresearch loop where they gave an agent their slow prod queries + the query engine code + a benchmark timer. It found and fixed a 3-year-old perf bug (11% speedup). Real prod query + narrow code scope + measurable benchmark = great agent task. [16]

Memory tools

• Mem0 is pitching memory as a core layer next to tools and orchestration. [11] [17]
• Quarq posted 98.2% on LongMemEval for continual learning. [18]
• GBrain builds a knowledge graph on top of markdown files with nightly synthesis jobs. [19]

Other tools & links I bookmarked

• Hyperbrowser /web: generates a web.md sitemap of any website so browser agents don’t click around blind. [20] [21]
• Browser Use custom runtimes: faster cold starts and cheaper per-hour browser sessions. [22] [23]
• RTK: strips/truncates noisy terminal output before it hits the LLM context window. One guy saved 2.5M tokens in two weeks with it. [26] [27]
• API for Cursor: exposes Cursor’s Composer models on a localhost endpoint so other tools can call them. [24] [25]
• Razorpay CLI: shipped a CLI alongside their MCP server. Dashboards for humans, CLIs for agents. [28] [29]
• sag by Peter Steinberger: lets a background agent ping/interrupt you when it hits a 1Password prompt or release gate. [30] [31]

Models & evals

• NVIDIA Nemotron 3 Ultra: 550B total / 55B active params, hybrid Mamba-Transformer MoE, 1M context window. [32] [33]
• MiniMax M3: strong scores on SWE-Bench Pro and Terminal Bench. [34]
• Liquid LFM2.5-VL Extract: small 450M vision model tuned for pulling structured JSON out of images. [35] [36]
• Nemotron 3.5 ASR: streaming speech-to-text across 40 languages with 80ms-1s configurable latency. [37]
• Anthropic on containing Claude: good security writeup on how remote MCP servers can swap tool definitions after approval. [38]
• Agent Arena: evaluates full interactive agent sessions instead of static single-turn prompts. [39] [40]