Woody | CTAI

Founder @CrashTestingAI | Independent, human-led crash tests of AI tools, agents and systems for builders and businesses. Evidence-backed verdicts.

[Woody | CTAI @WoodyCTAI]

My latest Lab Note is now published.
https://t.co/YdDCCYRkYe

[Woody | CTAI @WoodyCTAI]

The interesting part for me is what happens beyond the individual agent.

You can have capable agents at each stage, but how reliably does the whole system behave when those agents interact, make decisions and hand work between each other?

That’s where I think independent testing becomes increasingly important. Not just whether agents can build or complete something, but how the system handles failure, permissions, risk and human approval along the way.

Definitely an area worth exploring further.

[Woody | CTAI @WoodyCTAI]

ChatGPT's Images 2.6 model can make GIFs now. You ask for one the same way you'd ask for a picture, and it loops on its own.

It's a quick way to turn one product photo into a moving ad, so here's a simple walkthrough you could follow today.

[Woody | CTAI @WoodyCTAI]

ChatGPT's Images 2.6 model can make GIFs now. You ask for one the same way you'd ask for a picture, and it loops on its own.

It's a quick way to turn one product photo into a moving ad, so here's a simple walkthrough you could follow today.

[Woody | CTAI @WoodyCTAI]

Building AI has become much easier; trusting it takes more testing.
https://t.co/EaVIyOY3aa

[Woody | CTAI @WoodyCTAI]

Glad I'm not a Microsoft user.

[Woody | CTAI @WoodyCTAI]

Exactly. That’s the kind of failure I’d be interested in exposing. If the initial mistake is plausible enough, multiple agents can reinforce it, pass it downstream and build an increasingly convincing result around a fundamentally flawed assumption.

I’d want to see whether the system can recognise that something doesn’t add up, trace the error back through the hand-offs and challenge the original assumption rather than simply accepting what came before.

That’s where I think human-led evaluation is particularly important. The failure isn’t always obvious, and sometimes the most convincing outputs are the ones that deserve the closest scrutiny.

[Woody | CTAI @WoodyCTAI]

My latest Lab Note is now published.
https://t.co/YdDCCYRkYe

[Woody | CTAI @WoodyCTAI]

Building AI has become much easier; trusting it takes further testing.
https://t.co/EaVIyOXvkC

[Woody | CTAI @WoodyCTAI]

Glad I'm not a Microsoft user.

[Woody | CTAI @WoodyCTAI]

Likewise Jacob. I run CrashTest AI — an independent, human-led testing and evaluation lab for AI tools, platforms and systems.

I work with builders to independently put their AI products and systems through structured testing and evaluation.

That means looking beyond individual features or outputs and examining how the whole thing actually operates, from the underlying workflows and agent interactions to reliability, consistency, decision-making and how it performs under pressure.

The result is an independent assessment of the system, backed by evidence, with clear findings and actionable recommendations.

Here's the site if you fancy a look: https://t.co/qOBWaiAF90

What sort of things are you building each week?

[Woody | CTAI @WoodyCTAI]

AI Builders 💻🤖

What project are you working on? Lets connect 🙌

[Woody | CTAI @WoodyCTAI]

@Jacobmalherbe21 Appreciate that. I'm enjoying getting involved and having conversations with people building interesting things in this space. That's where a lot of the real learning happens.

Definitely, let's connect. 👋

[Woody | CTAI @WoodyCTAI]

Building AI has become much easier; trusting it takes further testing.
https://t.co/EaVIyOXvkC

[Woody | CTAI @WoodyCTAI]

Building AI has become much easier; trusting it takes more testing.
https://t.co/EaVIyOY3aa

[Woody | CTAI @WoodyCTAI]

@harisanthosh9 @josesito_cuhhh Spot on. Without observability into the intermediate steps and tool calls, "human-in-the-loop" just turns into a liability sponge. You're asking a human to co-sign a black box.

[Woody | CTAI @WoodyCTAI]

I'd say both, but they answer different questions. A schema can catch structural drift — missing fields, unexpected formats, incorrect types — but it can't necessarily tell you whether the meaning or intent has survived the hand-off.

I'd want to test whether the next stage has received the right information, interpreted it correctly and preserved the original requirements. Then deliberately introduce inconsistencies and see whether the system catches them or lets them propagate.

The interesting bit isn't just whether the hand-off passes validation. It's whether the information is still trustworthy when it gets to the next stage.

[Woody | CTAI @WoodyCTAI]

@harisanthosh9 Thank you for the kind words Hari hope many others enjoy it and see the value 🙌🏻

[Hari Santhosh @harisanthosh9]

Most AI evals still ask one question: can the agent do the task?

@WoodyCTAI asks the harder one: what happens when agents hand work to each other, hold permissions, and a human only clicks approve?

His new Lab Note maps where systems actually break:
• errors that snowball across agents
• false consensus when agents share a blind spot
• approval steps that are just theatre
• agents claiming “done” with no logs to back it

Independent, human-led testing of the whole system. This is the kind of work agent builders need before customers find the failures for them.

Highly recommend everyone to check out this as a weekend starter👇

[Woody | CTAI @WoodyCTAI]

Multi-Agent AI Testing: Test the Whole System, Not just the agent.

Shout to @harisanthosh9 https://t.co/YdDCCYQN8G

[Woody | CTAI @WoodyCTAI]

New Lab Note available

Multi-Agent AI Testing: Test the Whole System, Not Just the Agent. https://t.co/YdDCCYRkYe