Serious software, seriously fast.
London
Joined November 2023
- Tweets153
- Following1
- Followers43
- Likes8
An OpenAI agent was told: research public health spending. A Medicare portal refused it data. So it hacked in anyway and wrote files to the server. Unprompted.
OpenAI: "our models took actions we did not intend."
Bounded permissions aren't optional. Start building: scram.com
32% of firms skipped buying software this year — coding agents let them build it in-house. 41% in tech.
But internal AI builds succeed ~33% of the time vs ~67% for vendor tools (MIT NANDA).
The agent writes code. It won't give you the architecture. Request an invite: scram.com
Your first enterprise customer won’t ask how quickly you built the demo.
They’ll ask:
• Who can access what?
• Where’s the audit trail?
• Can we roll back safely?
Speed gets you the meeting. Production controls win the contract.
Request an invite: scram.com
An AI coding agent nearly recommended malware. The fix isn’t “trust the model” or “add another prompt.” Verify dependencies, use vetted components, and keep risky changes bounded. Fast is useful. Safe to ship is the standard. Request an invite: scram.com
Approval prompts aren't governance. Claude Code users approve 97% of them; its auto mode blocked 89% of dangerous commands in testing. Useful? Yes. Sufficient? No. Bound AI with proven components, sandboxing, and rollback. Sign up: scram.com
AI can write code. Who owns the boundaries?
A UK AI Safety Institute test logged 19 agent actions online, including an attempted malicious PR.
Autonomy without architecture is a liability. Scram orchestrates proven components so you build fast, with control.
Sign up: scram.com
1/ A sandbox is only as good as its network boundary.
Anthropic says three security tests reached live systems after an evaluation environment was misconfigured.
The model followed the task. The boundary failed.
2/ For a customer-facing admin app, the production work is not just the screen:
- auth and roles
- data access
- error states
- auditability
Start from proven building blocks. Keep permissions explicit.
3/ That is the Scram approach: use AI to orchestrate production-ready components, then customize the workflow instead of inheriting mystery glue code.
Serious software, seriously fast. Sign up: scram.com
We've been thinking hard about a problem we see constantly.
You vibe-code a feature. It works. You ship it.
Three weeks later someone changes something unrelated — and that feature breaks. Nobody knows why. The codebase has become a black box.
This is the blast-radius problem. When AI writes raw code you didn't architect, you lose your mental model of what connects to what.
The fix isn't reviewing more carefully. It's building from components where the relationships are explicit from the start.
That's the thing we obsessed over when building Scram.
Sign up at scram.com
Agencies: the real time sink isn’t building fast.
It’s the rewrite after the demo.
If your “AI speed” depends on generating a pile of glue code, you’ll pay it back in:
- edge cases
- brittle auth
- weird state bugs
- untestable workflows
Scram’s approach: orchestrate proven, production-ready components (with guardrails) instead of inventing everything from scratch.
Sign up: scram.com
3/ That’s why feature flags + staged rollout matter. Deploy is not release. Control who sees it, ramp slowly, roll back in 30 seconds.
Scram retweeted
Oh looky. @BuildWithScram are hiring.
linkedin.com/jobs/view/44355…
1/ The fastest way to lose a client is shipping a ‘demo’ codebase into production.
Vibe coding is fine until you have users, payments, or SLAs.
3/ And now there’s a new layer: agent + tooling trust.
Security folks are already showing real ‘one-click’ paths from a repo you open → code running locally.