11factor / agentic engineering

AI agents are easy to demo.Hard to depend on.

I’m Sandeep. I write about engineering with AI agents and building systems people can depend on. 11factor is my framework for the work: clear tasks, useful context, controlled execution and evidence before improvement.

Read the writingSee the projects

Latest writing

Notes from the work.

All writing →

I write about tasks and specifications, context, tools and orchestration, evaluations, observable execution and permissions. Short notes and experiments help me work out what improves the result.

27 Sept 2026Jev vs LLM-as-a-Judge: What I Found Testing Both on a Real AgentI ran our hand-written eval rules, TypeSafe's Jev and four frontier LLM judges side by side on 239 labelled cases from a real family-assistant agent. Here's where each one earns its place.9 min read
21 Sept 2026BERT, an LLM, or Jev: what should make the decision?Compare encoder classification, causal generation and Jev's bounded decision interface, while keeping its unknown internals explicit.17 min read

11factor / agentic engineering

Build. Run. Observe. Improve.

My framework connects the task to execution, evidence and the next change. Human control, permissions and version lineage apply throughout. I have not demonstrated the full end-to-end learning loop.

Working question

Can agents learn from real work without becoming less reliable?

Build

  1. 01Model access & selection→
  2. 02Context assembly→
  3. 03Memory & retention→
  4. 04Tools & data access→

Run

  1. 05Execution state & orchestration→
  2. 07Operational limits→

Observe

  1. 08Evaluation & observability→

Improve

  1. 09Self-learning & recursive improvement→

Improve feeds authorised changes and later outcomes back into future Build and Run attempts: repeat the cycle with fresh evidence.

Across the loop

Human intervention, externally enforced authority and reversible changes span every stage.

  1. 06Human control→
  2. 10Permissions & safety→
  3. 11Versioning, audit & rollback→
Read the framework →

Projects

Where the ideas get tested.

Oyako and Clinly are working projects. They show where I apply the thinking, with clear limits on what each project proves.

Oyako

An AI co-parent that turns school emails and family admin into dates, reminders and clear next actions.Public / pre-launch

Clinly

A two-sided marketplace for clinicians seeking expert case review and mentoring from specialist clinicians.Private / in development
See the projects and evidence limits →

My setup

How I organise work with agents.

This is the setup behind my agentic-engineering work. Mothership is how I organise profiles, project context and Kanban around Hermes. Hermes Agent, Claude Code and Codex are the tools I use for the work.

Task workHermes AgentI use Hermes Agent a lot when I want agents to work on a task.
Deeper testingClaude Code and CodexI often use Claude Code and Codex directly when I want to test something more deeply.
See my setup →

Sandeep Mehta

I’ve spent more than a decade in distributed systems. Now I’m exploring the engineering around agents.

About me