Directing AI agents
Brief an agent so it knows the outcome, what it may touch and when to stop, then check what it did before it ships.
First read: How to scope an AI agent task →AI consulting in the UAE
I help UAE teams brief, run and verify AI agents before anything goes live.
Will this model fit your machine?
20.3 GiB of 32
Estimated memory fit within the GPU budget
Weights 15.8 GiB, cache 2 GiB, runtime 2.4 GiB, with 64 GiB of system RAM assumed. Verify with a real request before relying on it.
Start with the job in front of you
The tools change every month. What lasts is knowing how to brief an agent, how to check what it did, and how to keep what you learned for the next run.
Running models on your own hardware
Which models fit which machine, measured on an M4 Max and an RTX 5090, with the calculator I open first.
First read: Will this model fit? →Your team wants one workflow checked
Ten working days on one recurring workflow, ending in a written decision: proceed, proceed with limits, or do not automate.
First read: the Verified AI Operations Sprint →27 notes · 70 guides · 31 templates
LatestClaude Fable 5.2: pricing, benchmarks and what changed
My guides document how I work with agents and local models, including the tests, failures and checks behind each recommendation.
Explore my methods →Tap a star to see the page it stands for.
Claude Fable 5.2: pricing, benchmarks and what changed
Claude Fable 5.2 costs $8 in and $40 out per million tokens, down from $10 and $50. Specs, scores against Opus 5.5, and what breaks on upgrade.
Changing the rules must not rewrite the past
A schedule change also changed what counted as a complete day of records. I dated the policy and kept earlier sessions under their original rules.
An engine update is not a model test
The LM Studio engines were updated and the SGLang service answered a real request afterwards. Generation through the new engines stayed untested.
Move the finish line to the outcome
A workflow reported success before the downstream work had finished. I moved the completion check to the destination.
Guides people come for
LLM VRAM calculator: will this model fit your GPU or Mac?
Estimate model weights, KV cache and runtime overhead for a GPU or Mac, then verify backend compatibility and peak memory with a representative request.
SGLang vs vLLM on an RTX 5090: speed, memory and context
SGLang and vLLM on an RTX 5090: generation probes with a configured 32K cap, separate long-prompt and stability tests, and measured memory use.
Local LLMs on Apple Silicon: which Mac, RAM and speed
Mac memory budgets, dated M4 Max decode measurements and context costs, with exact build, runtime and comparison limits kept beside the figures.
Running AI Agents in Production: Scope, Verify, Monitor
Four moves for production agents: scope each task, let agents act alone only where they can undo it, check outcomes not green ticks, record decisions.
The verification loop: five checks for AI-assisted work
Five checks before calling AI work done: name the claim, record a baseline, test the behaviour, scan what else it touches, and look at the rendered result.
AI agent handoff template: what the next session needs
My handoff packet answers five questions in order: what is true, what was decided, what proves it, what must stop the work, and the next safe action.
Dubai, and a desk with two machines
An M4 Max and an RTX 5090 run the local models these guides measure. The notebook started on 15 January 2026.
Work with me on one workflow
Start with one recurring workflow. The Verified AI Operations Sprint is a readiness assessment for UAE teams. Over ten working days after approved inputs are ready, I test one change outside production and give your team a written recommendation: proceed, proceed with limits, or do not automate.
FrameDays 1–2
Diagnose known unknownsDays 3–4
Create the loopDays 5–7
Build the workflowDays 8–9
DecideDay 10