The Great Transition Honestly, I don’t think the human brain is wired for what’s coming over the next 36 months. We are about to enter a period of profound cognitive dissonance that will be felt in every corner of the globe. https://t.co/MSEO64N7wZ

A practical AI system maturity ladder: Single prompt Prompt chains / DAGs Single agent + tools (RAG, function calls) — large gap — Multi-Agent System Advance only when observability, reliability, or expressiveness break at the prior layer.

The objective is correctness, reliability, and user value — not architectural novelty. Don’t trade debuggable code for opaque behaviour unless you must. If the task is “summarise a PDF and send an email,” a researcher → writer → emailer agent mesh is unnecessary, brittle, and

A practical AI system maturity ladder: Single prompt Prompt chains / DAGs Single agent + tools (RAG, function calls) — large gap — Multi-Agent System Advance only when observability, reliability, or expressiveness break at the prior layer.

MAS is the right tool in specific domains. Examples: • Market simulations with competing incentives • Multi-objective planning with negotiation • Decentralized systems with local knowledge If your agents don’t have conflicting goals or independent policies, you probably don’t

The real cost of MAS is loss of determinism. Multiple agents introduce: • Emergent behavior • Stochastic inter-agent dependencies • Non-reproducible failures Debugging shifts from reasoning about code paths to post-hoc behavior analysis. That’s a tax most teams underestimate.

We’re now applying the same thinking to AI systems. More agents ≠ more intelligence. More autonomy ≠ better outcomes. In many cases, it just means: • More state • More non-determinism • Fewer guarantees

This mirrors an older infra mistake. Kubernetes was adopted for small CRUD services because hyperscalers used it. The result: operational drag, not leverage. Powerful abstractions only pay off once failure modes, scale, and variability demand them.

Most workloads people apply MAS to today are structurally simple. They are solvable with: • Well-specified prompts • Deterministic orchestration • Explicit control flow (A → B → C) If agents only execute predefined handoffs, you haven’t built autonomy — you’ve built an

Unpopular opinion in the current AI hype cycle: You probably don’t need a Multi-Agent System (MAS). We’re at peak “architecture first, problem later.” Agent swarms are being proposed before teams can ship a single deterministic chain. Thread 👇

Embedded SQL is the silent killer of cloud migrations. Not in your schema inventory. Not in your runbook. But it will break cutover—or worse, change results quietly. Where it hides + how to smoke-test it early: https://t.co/j97bexRJIn via @smartmigrate

If I was starting an AI initiative today on GCP, my playbook would be: • pick one painful, measurable workflow • wire it properly into Vertex + BigQuery • obsess over logs/evals for 90 days • only then talk about “expanding the platform” Everything else is noise.

Finally, treat “agentic” as architecture, not marketing. State, memory, and observability are not optional: • what did the agent know? • what step failed? • can we deterministically replay? If you can’t answer those, you don’t have an agent. You have a black box.

Fifth, don’t try to replace the human immediately. Start with: • suggestions • draft RCAs • proposed fixes Then graduate to: • auto-resolve only in low-risk paths • human-in-the-loop everywhere else Trust is a migration path, not a toggle.

Fourth, bake evals in from day 1. For every action the system takes, log: • input • decision • human correction (if any) Once a week, replay the worst 20 and ask: “Should this system even be allowed to do this?” Most teams never do this review.

Third, over–invest in data wiring, not prompts. On GCP that means: clean slices of data exposed via BigQuery clear contracts on what the agent can’t touch one place where you log every decision it makes If the data is a mess, your prompts won’t save you.

Second, decide the system boundary early. Is this: a copilot for an SRE on call? an autonomous step inside an incident runbook? a read-only advisor? Scope creep is where AI projects quietly die.

First, stop building “demos for demo’s sake”. Every POC should answer one question: “Which existing metric will move if this works?” MTTR cost per ticket sales cycle time Pick one. Design backwards from that.

Most AI POCs in enterprises are dead on arrival. Not because the models are bad. Because nobody designs for “how does this survive contact with reality?” Here’s how I’d fix that, especially on GCP/Vertex:

SmartConvert = Precision at Scale for database migrations 🤖 95% automation via 1,000+ conversion rules + AI ⚡ 80% faster than manual approaches ✅ 100% accuracy with validation and reconciliation Oracle | SQL Server | Teradata → BigQuery | Redshift | Cloud SQL :

@jmrphy funny how the most interesting stuff happens when people stop using AI ‘properly’