Skip to content
Aleem.

Blog

Notes from the workbench.

Newsletter

New notes, in your inbox.

Get each new article from the workbench by email. No spam, unsubscribe anytime.

An always-on agent computer carries a recurring task through research, drafting, and evidence before returning it for human review.

Always-on AI agents: what OpenAI Dots actually changes

Always-on AI agents can carry work between conversations. Learn what OpenAI Dots changes, where agent computers help, and how to run a useful pilot.

 Four model output cards converge on a blue accepted-result lens beneath the headline The Task Chooses the Model.

Sprinting the Frontier: Reviewing and Comparing Opus 5.5, Sonnet 5.5, GPT-6.1 Sol, and GPT-6 Astra

Compare Opus 5.5, Sonnet 5.5, GPT-6.1 Sol, and GPT-6 Astra on real work, pricing, benchmark limits, and cost per accepted result.

 An approved app brief is cut into testable feature slices that converge on a verified live browser journey.

Claude Code in Production: Build Apps Without Vibe Coding

Use Claude Code in production with a one-page PRD, vertical task slices, test gates, and release checks. A practical app-build workflow.

 A task brief splits into GPT-6 Sol's compact route and Astra's multi-step route, both ending at an accepted-result gate.

GPT-6 Sol vs Astra: The Model Shootout That Matters

GPT-6 Sol vs Astra is a workload decision, not a leaderboard contest. Compare capability, speed, and cost, then run a fair shootout on your own tasks.

 A creative brief feeds game, animation, and audio parts into a finished interactive project with a human review checkpoint.

Claude Opus 5.5: What Creators Are Making

Claude Opus 5.5 is helping creators build games, videos, and interactive projects. Here is what changed, what examples show, and where review matters.

 A bounded AI improvement loop: a model proposes changes, tests evaluate them, and a human reviews what ships.

Gemini RSI Leak: What the Evidence Actually Proves

Gemini RSI leak claims are outrunning the evidence. See what Google has confirmed about recursive improvement, RRSI, and AlphaEvolve.

 Language flows into a central judgment gate, then routes to deterministic code, human review, or an uncertain-case path.

Jev Judgment Models Are Fast. That Does Not Make Them Better Than LLMs.

Jev judgment models make fast, typed decisions inside software. Learn where they beat LLMs, where they do not, and how to test them safely.

 A GPT-6 Astra creator workflow stays inside a defined operating boundary, with a human approval gate before public action.

GPT-6 Astra Is a Capability Leap. The Workflow Is Still the Product.

GPT-6 Astra can carry more work across browsing, computer use, and creation. The advantage now comes from supervised creator and agent workflows.

 A dark dashboard-style illustration showing a 12-week qualified-lead target converted into weekly milestones and a Monday review loop.

How I Turn SMART Marketing Goals Into Weekly Targets With Claude Code

Use Claude Code to turn vague marketing goals into a measurable weekly system. Set one outcome, choose the right metric, and review it every Monday.

 A reconciliation ledger showing that strategy survives delivery only when each decision is carried into a downstream artifact.

Marketing System Build: Keeping Strategy Alive Through Delivery

A case study of the 19-document marketing system I built to carry strategy, research, and brand rules from intake to published content.

 Automation is a single fixed rail with stops written in advance. An agent is a branching tree choosing its path at each node. Same start, different auditability.

AI agent vs automation: most of what you want is a scheduled job

Automation follows a path you wrote. An agent picks its own. Here is the test that tells you which you need, and why buying the wrong one costs more.

 A hundred-cell grid of skills, 88 filled, with a marked line at 51 showing where the last count stood, arguing that the system is not finished, it is compounding.

88 Skills, 194 Decisions Logged, One AI System That Actually Runs My Business

I stopped chatting with AI and built it an operating system: 88 skills, a second brain, and 194 logged decisions.

 Three concentric layers, the business, a review layer, and the automated task, with the middle review layer drawn as a dashed, incomplete ring to show what a small team has to build by hand that enterprise case studies already have in place.

Claude Code for Small Business: What the Enterprise Case Studies Don't Show You

Enterprise Claude Code case studies come from Datadog and Rakuten. Here's what actually changes when a small team runs it instead.

 A balance scale sitting level between Claude Code and Codex, because their SWE-bench Verified scores are now a single point apart.

Codex vs Claude Code: pick by task type, not the benchmark

Codex and Claude Code now score within a point of each other on SWE-bench. Three real-task tests show what actually decides it instead.

 A switchboard patching one finished brief through four sequential passes into three separate content channels, replacing mood-based, random posting with a routed system.

The 12-Question Marketing Brief That Ends Random Acts of Marketing

I stopped posting on mood and built a 12-question brief that turns one idea into a week of content in 2 to 4 hours.

 A skill fails visibly on one task. An agent's failure propagates across a chain of steps.

AI agent vs skill: I built 104 skills and zero agents

A skill is instructions for one task. An agent owns the loop. After building 104 skills and shipping few agents, here is when each one actually wins.

PROJECT IN MIND?

Let's build the system that runs it.

Describe your bottleneck and get a custom 2-minute Loom video showing exactly how to automate it, no call required.

Let's talk