Swamp Is Interesting Because It Doesn't Trust AI

Swamp Is Interesting Because It Doesn't Trust AI

Swamp Is Interesting Because It Doesn't Trust AI

Mark Ellens

Jun 12, 2026

•

2 min read

Most of the AI tooling world seems to be heading in the same direction. More autonomy, more agents, more black boxes.

The demos are impressive but the reality is often a collection of prompts, shell scripts and crossed fingers. So much has accreted and been bundled into a 'harness'

Which is why I find Swamp so interesting.

Not because it's the most powerful agent framework or because it promises AGI.

Not because it claims to replace engineers.

Actually, almost the opposite.

The Reliability Problem Nobody Talks About

I've spent the last few years working in platform engineering and SRE.

The most reliable deployment pipeline isn't the smartest one. It's the one that performs the same checks, in the same order, every single time. It's the one situation where you want a box ticking exercise to make sure that everythign in the environment and the build is as you want and expect it to be.

Every Time Claude Writes a Bash Script...

One of my running jokes recently has been:

❝

Every time Claude writes a shitty bash script, I think: this should be in Swamp.

What I mean is that the logic I'm interested in isn't usually the implementation, it's the workflow.

I don't want to think about:

shell scripts

temporary files

orchestration glue

process management

I want to think about:

what checks should happen

in what order

what evidence gets collected

what should happen if something fails

That's a workflow problem - Swamp feels like it's attacking workflows as a first-class concern.

The Most Interesting Thing About Swamp Isn't AI

The most interesting thing about Swamp might not actually be AI.

It might be that it gives us a way to define and execute organisational workflows.

Go into any large organisation and you'll find people reinventing the same processes:

Ticket → Design → PR → Review → Deploy

Incident → Investigation → Validation → Fix

Change Request → Approval → Release

Security Finding → Assessment → Remediation

The question isn't whether AI can perform the work.

The question is whether we can finally describe the work.

Bigger Organisations May Need More Batteries

Patrick Debois made a good point recently that larger organisations will probably need more batteries included.

I think that might be true.

But I also think Adam Jacob's response was important.

The best workflows don't emerge because somebody predicts them centrally. They emerge because people solve real problems at the edges and patterns gradually become shared.

That's how Chef happened.

That's how DevOps happened.

And it might be how organisational AI workflows happen too.

Why Platform Engineers Should Care

Platform engineers have spent years building paved roads.

Golden paths.

Templates.

Guardrails.

Swamp feels like a natural extension of that thinking - what if we could define how an organisation investigates incidents, reviews changes, validates releases or assesses risk?

Not as documentation nobody reads.

Not as Confluence pages nobody updates.

But as executable workflows.

The Part I Find Most Exciting

The thing that really caught my attention isn't that Swamp can run workflows.

It's that it gives us a place to inject determinism.

I don't necessarily want a more autonomous system.

I want a system that reliably performs the checks I care about.

For someone coming from an SRE and platform engineering background, that feels like a much more interesting problem.

Maybe the future isn't autonomous agents.

Maybe it's reliable ones.

Keep Reading

© 2026 Reliability & Validation Engineering.

Powered by beehiiv