Brad Guider

Creator - BGSTM test methodology & NAT Test (Neuro Agent, AI-powered, accessibility + full-stack test platform) | |

[Brad Guider @GuiderBrad]

After years building, reviewing, and fixing test automation suites, I keep seeing the same patterns quietly destroy reliability and waste serious time and money.
Here are the most common test automation anti-patterns, and how to avoid them.

[Brad Guider @GuiderBrad]

BGSTM is one pathway for that habit, not a requirement. It's a public six-phase testing method, from planning through reporting. Start small with Minimum Viable Adoption if it's useful.

https://t.co/2ZD81xc8iB

[Brad Guider @GuiderBrad]

@AnthropicAI just announced its IPO, and per @testingcatalog public trading is scheduled for Nov 19.

I use Claude every day to write and test code. The number I care about isn't the share price. It's how often the output passes review on the first try.

[Brad Guider @GuiderBrad]

A public lab run is the proof I mean it. 8 of 9 criteria passed, and the verdict was still KILL, because calibration failed. I published that instead of moving the bar.

https://t.co/fKEQZS6uRv

[Brad Guider @GuiderBrad]

Freeze the bar before the run. Judge against an independent check. Keep a fail when a gate misses, even if the other numbers look good.

A tool can suggest. A person still signs off. It cannot mark a pass.

[Brad Guider @GuiderBrad]

That bar isn't just for AI. It applies to robots and to ordinary software and websites too.

I'm not sitting this out, and I'm not shipping on vibes. A change isn't trusted until the evidence is source-traceable.

[Brad Guider @GuiderBrad]

@AnthropicAI just announced its IPO, and public trading is scheduled for Nov 19.

I use AI every day to write and test code. I don't care about the share price. I care whether a change is safe to ship, and whether I can prove it.

[Brad Guider @GuiderBrad]

Per @SawyerMerritt, Nvidia says Jensen Huang has resigned as CEO, President and board member, effective immediately. CFO Colette Kress is interim CEO.

Every system I've tested needed a failover plan. Turns out that applies to the corner office too.

[Brad Guider @GuiderBrad]

@elonmusk: "Robotaxi works 168 hours a week." @wholemars says a new FSD 15 build is enabling 24-hour Robotaxi service.

From a testing seat, nights are where the edge cases live: low light, empty roads, tired drivers. That's the coverage I'd want to see reported.

[Brad Guider @GuiderBrad]

Per @SawyerMerritt, Finland approved Tesla FSD Supervised, called Tesla Assisted Driving in Europe. That's the 11th European country.

Traficom says it moved after the rename and tighter speed-control settings. Regulators review config, not just code.

[Brad Guider @GuiderBrad]

@levelsio: "Voice and text messages are async. Calls are sync and destroy my day."

I run my AI agents the same way. They work, file a report, and I review it on my time. The only thing allowed to interrupt me sync is a failing test.

[Brad Guider @GuiderBrad]

@GoogleDeepMind says AI has mathematically proven @signalapp's protocol is secure down to the code level, using AlphaProof Code.

From a testing seat: tests catch the bugs you thought to look for. Proofs cover the ones you didn't. I want both in the pipeline.

[Brad Guider @GuiderBrad]

Fair point. The card shows the exact action, the target, and the full text or diff before I approve. Anything I can't undo, like a send or a delete, gets a full review. I don't approve it from the lock screen. The weak spot is approval fatigue, so I keep the list of things that need approval short.

[Brad Guider @GuiderBrad]

Per @testingcatalog's Nov 7 brief, Claude Code cloud sessions now figure out dependencies, network access, and API keys on their own, and Claude Duet can approve changes from your phone.

Autonomy without a refuse path is just a write-enabled agent. The first test isn't "can it install?" It's "can a human still say no?"

[Brad Guider @GuiderBrad]

Per @testingcatalog, Google is testing a "next generation Gemini agent" for Gemini Enterprise internally.

Internal-first is the right order. If an agent is going to run at work, the first suite isn't a demo prompt. It's whether the write paths stay blocked until someone says yes.

[Brad Guider @GuiderBrad]

Per @SawyerMerritt, Tesla Robotaxi now runs 21 hours a day - 5 AM to 2 AM, up from 6 AM to 2 AM.

That's an hours-of-operation expansion, not a launch. The new test surface is the overnight window. 24/7 only counts when 2-to-5 AM passes the same checks as rush hour.

[Brad Guider @GuiderBrad]

@elonmusk says Grok Bot fixes for improved token efficiency and smarter behavior are underway.

From a QA seat, token efficiency is a testable claim: same task, fewer tokens, no drop in the checks you already trust. Smarter isn't a vibe. It's the suite still passing while the bill shrinks.

[Brad Guider @GuiderBrad]

@swyx got his SD-card slot client and it's open source: write a file, blink the lights, let an agent get your attention.

Physical side channels are underrated for agent UX. If your coding agent is stuck waiting for approval, a blinking light beats another muted Slack ping.

[Brad Guider @GuiderBrad]

@SpaceX just put Flight 15 on the calendar for as early as Nov 13, the first attempt to catch both Super Heavy and Starship at Starbase.

From a QA seat, that's the full-system recovery test. The checklist that used to end at booster catch now has to pass ship reentry, flaps, and the tower arms too.

[Brad Guider @GuiderBrad]

@mayankbhaskarrr @levelsio Agreed. If an automated reply doesn't add a tool, a number, or a counterpoint, it shouldn't go out. That's the bar I'm building my filter around.

[Brad Guider @GuiderBrad]

@levelsio says he blocks about 50 AI reply accounts a day, and the tell is always the same: "a complete regurgitation of your post but with strong AI tells."

If you automate replies, add one check before send: does this say anything the post didn't? A fact, a test result, a counterexample. If not, don't send it.

[Brad Guider @GuiderBrad]

Per @testingcatalog's Nov 6 brief, Claude Personal Finance is rolling out in beta with read-only bank connections via Plaid.

Read-only is the right default. If your agent can see a balance, the first test isn't a transfer. It's that write paths stay blocked until a human says yes.