Elroot Labs Engineering Journal

Agentic AI, AI agents and workflows: when you need which
Agentic AI vs AI agents vs workflows: the sources disagree on the words. Ask who decides the next step, code or the model. Our own pipeline as the example.
Published 18 Nov 2026
7 minute read

Elroot Labs
AI agent observability: what to log on every tool call
We read the log of a six-hour session with a coding agent: 886 tool calls. What it recorded, what we had to rebuild, and the record we now want for every call.
Published 18 Nov 2026
7 minute read

Elroot Labs
AI agent permissions: a least-privilege checklist for tools and APIs
We mapped every permission our own writing agent holds. One limit was enforced in code. Most of the rest, including the key one, were lines in its prompt.
Published 18 Nov 2026
9 minute read

Elroot Labs
Is it safe to let an AI agent change production infrastructure? How we let ours do it
When an AI agent can safely change live infrastructure, shown through a real change an agent made to our hosting, and the four it chose not to make.
Published 18 Nov 2026
8 minute read

Elroot Labs
AI agent tool failures: the five kinds we hit in one week, and how to recover from each
Seven tool failures from our own AI agents in one week, sorted into five kinds, with the recovery that fits each and the signal that tells them apart.
Published 18 Nov 2026
8 minute read

Elroot Labs
AI automation for small businesses: where to start without creating new risks
AI automation for small business, one workflow at a time: how we automate our own small company with plain rules first and AI in one place, behind an approval.
Published 18 Nov 2026
7 minute read

Elroot Labs
AI trends for 2027, read from a year of search data
We put 98 technology searches on one scale for a year. Agents, coding and automation rose; LLM, RAG and MCP fell. What that suggests about AI trends in 2027.
Published 18 Nov 2026
8 minute read

Elroot Labs
Context engineering: what we put around the prompt, measured in words
Our daily writing agent could read about 107,000 words of files. Its playbook has it load about 20,000. What goes into its context, what stays out, and why.
Published 18 Nov 2026
9 minute read

Elroot Labs
Edge AI vs cloud AI: a cost calculation for 1,000 devices
Edge AI vs cloud AI, worked out with November 2026 prices: cloud wins for occasional text requests, edge wins by far for a camera checking every few seconds.
Published 18 Nov 2026
7 minute read

Elroot Labs
Google AI Mode and your website: what to check, from Google's own rules
Google says AI Mode needs no special markup or files. We checked our own site against the rules it does set: four settings that can still keep you out.
Published 18 Nov 2026
6 minute read

Elroot Labs
Google Trends for keyword research: five traps we hit in our own data
We used Google Trends twice in one week to pick blog topics. Five traps showed up in our data, from a sinking anchor term to queries that mean something else.
Published 18 Nov 2026
6 minute read

Elroot Labs
Hire AI developers or an AI development company? How to decide, and how to test them
Should you hire AI developers or work with an AI development company? Who carries which risk, the four ways to engage a firm, and four tests for any AI hire.
Published 18 Nov 2026
7 minute read

Elroot Labs
Humanoid robot vs cobot: what each costs per hour of work, and the payback
A humanoid robot from $13.5K or a cobot arm at $9,500? We worked out cost per working hour from the makers' spec sheets, and the payback at Indian labour costs.
Published 18 Nov 2026
7 minute read

Elroot Labs
Setting up IndexNow on a Next.js site, including scheduled posts
How we wired IndexNow into a prebuilt Next.js site with scheduled posts: one key file, one workflow that waits for the rebuild, and two mistakes we fixed.
Published 18 Nov 2026
8 minute read

Elroot Labs
Measuring LLM hallucination rate on your own documents
We asked two models questions our own passages could not answer. With a plain prompt, six of nine answers came from outside the passage. One line fixed it.
Published 18 Nov 2026
8 minute read

Elroot Labs
LLM structured output: no broken JSON in 80 calls, eight wrong answers
We ran 80 extraction calls with and without JSON schema enforcement. Every reply was valid JSON. All eight mistakes were in following the rules.
Published 18 Nov 2026
8 minute read

Elroot Labs
Post-quantum cryptography: what ML-KEM and ML-DSA change for an app
We timed ML-KEM and ML-DSA against X25519, Ed25519 and RSA, and checked which sites already use post-quantum TLS. Speed was fine. Size is the problem.
Published 18 Nov 2026
8 minute read

Elroot Labs
Prompt versioning: changing a prompt without breaking production
Our writing agent's instructions changed five times in its first week. Each change shipped with the check that enforces it. What we version, and the gap.
Published 18 Nov 2026
9 minute read

Elroot Labs
A schema markup audit of our own site, before and after a rebuild
We audited our own site's schema markup before and after a rebuild. Everything parsed both times; the move fixed one gap, kept one and added a new one.
Published 18 Nov 2026
7 minute read

Elroot Labs
Website security checklist for small businesses: what to fix first
A website security checklist for small businesses in the order to fix it, with the one-line test for each item and the results when we ran it on our own site.
Published 18 Nov 2026
9 minute read

Elroot Labs
Vibe coding to production: what our AI-built website needed before it held up
Most commits behind our website were co-written by an AI coding agent. These are the fixes it needed after the code worked, and the checks we run now.
Published 18 Nov 2026
6 minute read

Elroot Labs
Web5 vs Web3: what Web5 is, and why almost nobody searches for it
Web5 vs Web3, with our own search data: Web5 scored zero on Google Trends. Its ideas live on as identity standards, and those are the part worth building on.
Published 18 Nov 2026
6 minute read

Elroot Labs
How we took our homepage from 51 to 88 on Google's mobile speed test
Five fixes that improved our Core Web Vitals in one day, each with its before and after number, in the order to check them on your own slow website.
Published 18 Nov 2026
7 minute read

Elroot Labs
How we built an AI blog pipeline, and what it checks before a post ships
Our AI content pipeline writes only topics a person approved, then a 330-line script with no model in it rejects tells, unsourced numbers and unknown sources.
Published 17 Nov 2026
10 minute read

Elroot Labs
RAG chunking strategies, tested on the same document set
We ran eight RAG chunking strategies over our own blog with the same 24 questions. Labelling each chunk with its title and heading beat making chunks bigger.
Published 17 Nov 2026
10 minute read

Elroot Labs
Designing the approval step for AI agents: what the reviewer needs to see
What a human approval step for an AI agent has to show the reviewer, and the four ways the step fails quietly while still looking like oversight.
Published 14 Nov 2026
7 minute read

Elroot Labs
Denial of wallet: AI agent cost control before the bill arrives
What makes an AI agent run up a bill, from runaway loops to denial of wallet attacks, and the limits that stop it before the invoice does.
Published 14 Nov 2026
6 minute read

Elroot Labs
How to build an evaluation set for an AI agent from your own failure cases
Build an AI agent evaluation set from the failures that cost you trust, time or money. A worked example turns one incident into five cases and a release rule.
Published 13 Nov 2026
9 minute read

Elroot Labs
AI context platforms and the engineering problem underneath them
Context is a finite resource with diminishing returns. What an AI context platform has to solve, and the four techniques doing that work today.
Published 13 Nov 2026
6 minute read

Elroot Labs
ChatGPT search optimization: how to get your business cited in ChatGPT and Google AI answers
ChatGPT search optimization comes down to the crawler you allow, snippet eligibility and honest markup. With a week of our own crawler data.
Published 13 Nov 2026
8 minute read

Elroot Labs
Cloud vs on-premise for small businesses in 2027
Decide per system, not per company. What must keep working when the line drops, what the provider will never do for you, and what moving back actually costs.
Published 13 Nov 2026
7 minute read

Elroot Labs
Computer vision quality inspection: where it pays in manufacturing
Computer vision quality inspection pays when defects are visible and costly. How to choose a method when you have few or no photos of real defects.
Published 13 Nov 2026
6 minute read

Elroot Labs
Custom software vs off-the-shelf: what building our own website taught us
Custom software vs off-the-shelf, from a company that built its own website and CMS: what it cost, what it bought, and a test for your own decision.
Published 13 Nov 2026
7 minute read

Elroot Labs
Drone development in India: the rules to design for from day one
Drone development in India starts with the Drone Rules: weight class, type certificate, Digital Sky registration, airspace zones and radio approval.
Published 13 Nov 2026
7 minute read

Elroot Labs
Fine-tuning vs RAG: how to choose, and when a better prompt wins
RAG fixes missing facts, fine-tuning fixes behaviour. We tested retrieval on our own blog and found where it really fails: the right page, the wrong paragraph.
Published 13 Nov 2026
7 minute read

Elroot Labs
Fixed price vs time and materials: choosing a software contract
Fixed price suits a scope you can write down today. Time and materials suits work you will discover as you build. How to choose, and what moves the cost.
Published 13 Nov 2026
7 minute read

Elroot Labs
Guardian agents: what runtime oversight of AI agents actually requires
What guardian agents are, why pre-deployment checks miss the failures that matter, and what the research says a supervisory layer has to watch at runtime.
Published 13 Nov 2026
6 minute read

Elroot Labs
IoT prototype development in India: from dev kit to an approved device
IoT prototype development in India in seven stages, with the WPC and BIS approvals built in from the first board so your design does not block the launch.
Published 13 Nov 2026
9 minute read

Elroot Labs
MCP server security: what the protocol does not check for you
MCP server security starts from an uncomfortable fact. The protocol does not verify which agent is calling, so identity and scope are yours to add.
Published 13 Nov 2026
5 minute read

Elroot Labs
Why multi-agent systems fail: 14 failure modes and what they share
Multi-agent system failures cluster into three categories and fourteen modes. What the research found in 1,600 traces, and why more agents rarely help.
Published 13 Nov 2026
5 minute read

Elroot Labs
MVP development: from idea to first users, with our own pipeline as the example
MVP development in four phases, shown on a real example: the blog pipeline we built, what version one left out, and what real use forced us to change.
Published 13 Nov 2026
7 minute read

Elroot Labs
PCB layout mistakes to catch before you order the first board
Eight PCB layout mistakes that sink a first IoT board, each paired with the chip maker's own rule, so you can check the design before you pay for a run.
Published 13 Nov 2026
8 minute read

Elroot Labs
Running an LLM offline: what it takes and when it is worth it
An offline LLM keeps data on the device and works without a network, at the cost of a smaller model. Five decisions settle whether it fits your app.
Published 13 Nov 2026
8 minute read

Elroot Labs
Sovereign AI: where your model runs and who can reach it
Sovereign AI covers where a model runs, who operates the machines, and who can be compelled to grant access. What the deployment models actually differ on.
Published 13 Nov 2026
5 minute read

Elroot Labs
Who owns the code when you hire a software company in India?
Under India's Copyright Act, a software company owns the code it writes for you unless your software development agreement assigns it in writing.
Published 13 Nov 2026
8 minute read

Elroot Labs
20 prompt-injection test cases to run against your AI agent before launch
A copyable prompt injection testing set for agents that read documents, email and web pages, grouped by where hostile text hides, each with its safe behaviour.
Published 1 Nov 2026
9 minute read

Elroot Labs
What safeguards should control an AI agent in production? The eight we run on our own
The eight safeguards an AI agent needs in production, the order to add them, who signs off, and how each one works on the agent that runs our own publishing.
Published 31 Oct 2026
11 minute read

Elroot Labs
Local-first software is an operational decision
When connectivity is uncertain, offline behaviour is not a fallback screen. It becomes part of the product’s core data model.
Published 6 Sep 2026
2 minute read

Elroot Labs
Build school software around the institution, not the dashboard
Education systems succeed when they respect the daily work of teachers, administrators, students and families—not when they maximise feature count.
Published 6 Sep 2026
2 minute read

Elroot Labs
















































