Engineering notes
that stay practical
Shipping lessons, build playbooks, and engineering insights from the LFG team.
Automate your engineering workflows
Drop your email, we'll reach out with ideas for your stack.
The Smallest Companies May Become the Biggest Producers
For most of software history, output scaled with headcount. AI breaks that relationship — and the manager of tomorrow may supervise 200 agents rather than 10 people.
Software Factories
\"AI coding agent\" will eventually feel like a transitional term. The next important abstraction above coding agents is not a better coding agent — it is the factory that coordinates them.
Models Become Commodities
The model may become one of the least interesting parts of an AI product. The model is intelligence; the harness turns intelligence into useful work — and that is where the durable advantage sits.
Open-Source Models Change the Economics
When a human chats with a model for a few minutes, inference cost is invisible. When an agent works for three hours and spawns twenty more, intelligence becomes a line item.
The Cost of Intelligence Is Falling
For centuries intelligence was expensive because intelligence meant people. AI turns portions of it into compute — which makes it behave less like labour and more like infrastructure.
India + Software Services
Indian IT won because it built an enormous system for turning technical talent into reliable software delivery. The next opportunity may be building that system again — with a workforce that includes millions of agents.
Why Your AI Agents Need Sandboxes (And How Firecracker Makes It Possible)
AI agents are getting more capable by the week. But do you really want an autonomous agent running loose on your machine? Sandboxes are the answer—here's how Firecracker makes it possible.
Background Agents: The Infrastructure Layer Your AI Workflow Is Missing
Most AI coding tools wait for you to prompt them. Background agents don't. They sit in your infrastructure, watch for triggers, and execute — while you're doing something else. Here's what that actually requires.
Claude, GPT, and Grok: Why LFG Supports Multiple AI Providers
LFG works with Anthropic, OpenAI, and xAI models. Here's why model flexibility matters for real product work — and how to pick the right model for the right task.
Implementing Mags Sandbox environment on Baremetal machines
We've been building Mags, a sandbox platform for AI agents. Each sandbox is a Firecracker microVM that boots in ~300ms from a pre-warmed pool. Workspaces persist via JuiceFS + OverlayFS — writes go to local disk for speed, then sync to S3 on completion. Our newest trick: instead of tearing down the VM after a job, we park it with the workspace still mounted. If the same workspace runs again within 5 minutes, we skip the entire mount sequence and go from ~4s to ~100ms. Two 64GB bare-metal agents give us ~100 concurrent VMs.
Projects, Tickets, and AI: A New Model for Engineering Work
LFG replaces the traditional ticket tracker with an AI-native project model — where tickets aren't just tracked, they're executed by specialist agents against explicit acceptance criteria.
Real-Time AI: Why Streaming Responses Change How You Work
Watching AI output arrive in real time isn't just a UX nicety — it changes how you catch problems, course-correct, and collaborate with agents on complex work.
Why Every AI Code Action Runs in an Isolated Sandbox
LFG runs every ticket in its own Firecracker microVM. The AI can be aggressive and experimental — if something breaks, it breaks in the sandbox. Your codebase stays clean.
Transparent Tickets: Why Every AI Action Should Be Reviewable
AI that works autonomously is only trustworthy if you can see exactly what it did and why. LFG's ticket system makes every agent action visible, reviewable, and correctable.
How LFG's Orchestrator Turns a Single Request into a Multi-Agent Pipeline
Most AI tools give you one agent. LFG gives you a coordinator that reads your request, builds an execution plan, and dispatches specialized agents to work in parallel — while you watch.
AI First Software Agency
An enquiry into how an AI-first software agency is different — and why transparent, orchestrated AI execution beats prompt-and-pray.