Writing
Practical accounts of what I built, what made the work difficult, and what I would do differently next time.
46 articles
Coding agents generate enormous context — sessions, deliverables, files, decisions. The bottleneck isn't generating more; it's finding what you already made. Here's the small glue layer I built to solve exactly that, and why grep, FTS, and RAG each miss part of the problem.
Almost every browser automation tool claims to be the fastest, the cheapest or the most reliable. I put 34 of them through the same job with the same browser and one independent pass test, then ranked them twice — because speed and real work do not agree.
My assistant had 343 instruction sheets and could only see a fraction of them. The plain-English version of how I found the gap, why the built-in fix did not help, and what actually changed.
My coding agent used to spend extra model turns checking whether tests and deployment had finished. I moved that post-edit work into GitHub Actions on a self-hosted runner. The authoring model now commits, emits CI_HANDOFF and stops.
How I use ChatGPT as the controller, a private GitHub control repo and an Actions-dispatched Mac runner as the delivery path, and OMP/GLM as the cheap execution worker — including what broke in the first real end-to-end test.
The verified remote coding workflow I now run end to end: ChatGPT plans and reviews, GitHub holds the durable task state, Remote Desktop Commander bridges to the Mac, and a local OMP worker on GLM-5.3-Flash does the implementation.
A decision playbook for running Codex subagents without token cost surprises — when to delegate, how to bound each worker, and how to measure what it actually cost. Built from three measured experiments on this site.
The first AI use case worth doing in a traditional business, how to pick it, and what "clear business value" actually looks like — with a real shipped example and the honest failure modes.
Should you run server-side tagging? A real Cloud Run cost autopsy from rajeevg.com plus a plain decision framework: when same-origin sGTM pays for itself, and when client-side GTM is the better call.
How native Codex task controls, strict controller policies, persistent GLM workers and a CLIProxyAPI-to-OpenRouter route moved 92% of one PRD run's uncached prompt volume onto the cheaper model.
A plain-English guide for anyone who has an idea and a coding agent but no infrastructure experience — jargon explained, tokens stored safely, AGENTS.md doing the operational lifting, and a vendor-neutral way to ship a live website.
A practical guide to how one manager-owned Google Ads conversion setup can serve several country accounts and websites without the website deciding which account receives credit.
New AI models arrive constantly. This dashboard brings quality and price evidence together so it takes minutes, not hours, to find the models worth testing.
The full first-party measurement stack — consent, browser and server-side GTM, collection endpoints, BigQuery modelling, quality checks, privacy and activation — with an honest account of when the heavy stack is the wrong answer.
Files, email, decks, sheets and meetings become a governed markdown layer with named owners, review dates and provenance — so agents retrieve answers that trace to sources instead of folklore.
The chronological, evidence-labelled history of one rig's multi-agent architecture — what each layer promised, what the traces said, what was kept, and what was deliberately deleted — with the current recommendation stated plainly.
The portfolio moved from a flat projects list to a grouped solutions model with evidence dates, controlled links, and a separate record for earlier experiments. Here is what the inventory actually contains and why the framing matters if you are hiring, contracting, or buying.
A five-question decision rule — ambiguity, repetition, failure cost, real work, auditability — that decides whether a task deserves an agent, a workflow, plain code, or a human decision, worked through real examples from this rig and anonymised agency operations.
The measured mechanisms behind agent bloat: context resends, oversight cost, cumulative failures, and the attribution traps that make dashboards lie about all of it.
Ports, processes, services, containers, and one honest boundary: how a laptop becomes an operations centre you can reach from your phone without exposing anything to the internet.
A candid case study of the exact execution that shipped this site's solutions/editorial rebuild and visual remediation: a Sol supervisor, a GLM workhorse in an isolated worktree, a bounded Opus review, GLM amendments, and independent parent acceptance — with the failures included.
A beginner-friendly but technically complete tour of Rajeev's open-source agent telemetry stack: what each piece is, why it exists, where one real request travels, and what the stack still cannot prove.
A non-technical guide to using capable low-cost AI models for routine work, escalating the decisions that matter, and gradually automating the hand-offs.
What it actually takes to keep a fleet of AI agents doing real work — the contract-first operating loop, the failure modes that eat it, and a checklist you can run tomorrow morning.
How web analytics implementation — requirements, tagging plans, GTM and GA4 configuration, QA, governance and reporting — can now be substantially automated by agents, and exactly where humans keep control. Marketer-first, proof-backed.
Seven AI workflows run inside real agency operations — briefing, taxonomy governance, QA, evidence packs, reporting, knowledge retrieval, handoffs — with honest limits, human gates, and a map of where the machine earns and where it must stop.
How I connected Codex, OpenClaw, OMP and Hermes through the Linux Foundation A2A protocol, proved all 12 routes, and arrived at a much simpler coordination rule.
A delivery standard for AI-assisted work: raw evidence, provenance, editable deliverables, reconciliation, and review gates — walked through one real build end to end, including the failure that only the evidence caught.
What changed when I separated personal, work and general-purpose AI agents, then made their health, token use, cost and recent activity visible in one private Grafana control room.
A small OpenRouter routing benchmark made long GLM-5.3-Flash generations 63% faster, but modelling the result against a real day of agent use made the cost trade-off much clearer.
I ran four layers of orchestration between a request and actual work. The telemetry said most of it was overhead. Here is the before, the after, and the numbers that convinced me.
I pointed three different AI apps at one local gateway and one OpenRouter account, then wired their usage into self-hosted Langfuse and Grafana. Here is the full path, what it costs to run, and what the first real numbers show.
How I publish live Codex and OpenCode usage without publishing prompts, conversations, task names, or private telemetry systems.
How I kept Codex for planning and review, moved bounded execution to OpenCode Go, and measured the result without pretending I had proved savings.
An agent-readiness audit scored rajeevg.com 69 first and 75 in its refreshed report. Here is what changed, which gaps matter, and why I am not building an API just to chase 100.
How I deployed OpenClaw for research and Hermes for engineering as separate agents on a single Hetzner server, coordinated them through authenticated Agent Mail, and wired the result to Vercel deployment.
I added privacy-conscious session analytics to two agents running on Coolify, found broken Agent Mail calls and wasteful inbox pollers, then rebuilt the idle path to use zero provider tokens.
How I rebuilt my Codex subagent setup around reliable browser control, bounded work, recoverable handoffs, and honest token measurement.
I spent weeks recording how an AI assistant worked on my computer. The real lesson was not about cleverness. It was about repeating myself, checking work it said was finished, and finding a simpler way to work together.
A plain-English look at how I split Codex work between an OpenAI lead and lower-cost OpenRouter specialists, what the evidence shows, and what it does not.
A beginner-friendly account of how I connected a local Qwen vision model to Codex, added it to the supported model selector, kept web and image tools working, and fixed a context-breaking image bug.
A deep, quantified explanation of how browser privacy controls, consent choices, modeled analytics, warehouse gaps, back-end truth, and offline events can all pull marketing measurement in different directions.
A historical walkthrough of the March 2026 server-side analytics build on rajeevg.com, before the live site was simplified back to client-side GTM and GA4 only.
A detailed engineering walkthrough of the hackathon voting app, from single-screen UX and event-day safeguards to production proof, analytics, and the agent workflow used to ship it.
A granular walkthrough of the rajeevg.com GA4 property setup, captured during the March 2026 server-side stack pass and kept as a historical property-configuration record.
A visual, beginner-friendly guide to cloning a repo, working on a branch, opening a pull request, merging safely, and dealing with simple conflicts.