The Carbon Layer

Carbon-based humans, in a silicon world.

August 2026No. 10

A meta-harness has to derive the evidence it acts on

An agent told my prototype it had committed a fix, the commit didn't exist, and the only thing that caught it was code that asked Git instead of asking the agent.

19 min read▸ Video edition
July 2026No. 09

A meta-harness is coordination that lives outside the agents

I run four agent sessions to get one change reviewed properly and the coordination between them is still done by hand; writing down what that glue actually does produced eight components, none of which belong in a supervisor model's context.

16 min read▸ Video edition
July 2026No. 08

A domain agent is narrow enough to measure

The same agent answered one question straight from the database and confidently invented the next; a 25-question eval harness is how I found out which failures matter and what to fix first.

15 min read▸ Video edition
July 2026No. 07

The harness can improve without touching the model

An outside model found a real bug in my agent's harness, wrote a fix, and proved it helped; every safety property in that sentence had to be built by hand.

15 min read▸ Video edition
July 2026No. 06

A working coding agent is about 2,500 lines of harness

Fifteen chapters after the empty file, the model's weights hadn't changed once; everything that reads as agentic came from code we wrote around one API call.

10 min read▸ Video edition
July 2026No. 05

Context management decides what the model sees this turn

Forty turns into a real bug the agent started contradicting itself; the model never changed, the window did.

17 min read▸ Video edition
July 2026No. 04

Agent skills are procedures the harness can load

Your strongest reviewer cannot sit inside every pull request; a skill captures the repeatable part of their procedure so the agent stops rediscovering the job.

14 min read▸ Video edition
July 2026No. 03

Agent sandboxing is a blast radius decision

The word sandbox covers five very different mechanisms, from a V8 isolate to a microVM, and the right one depends on the failure you need to contain.

14 min read▸ Video edition
July 2026No. 02

Agent memory is context assembly over time

The model forgets everything at the session boundary; whatever feels like memory is the harness deciding what to keep, recall, update, and forget.

11 min read▸ Video edition
July 2026No. 01

Harness engineering is the system around the model

When an agent fails, the model takes the blame; the more useful questions are about the ten layers wrapped around it.

10 min read▸ Video edition
Ankit Desai
About

I'm Ankit Desai. By day I'm a Senior Engineer and AI Strategist. I've spent twenty-plus years on the layers that keep software running: cloud platforms, SaaS backends, deployment automation, reliability engineering. The last few years pulled me into agentic AI: developer tooling, evaluation infrastructure, governance, and what it takes for engineering teams to adopt AI workflows without lowering their quality bar.

The Carbon Layer is where I think out loud about that shift. Every post starts with something I actually ran: an agent, a harness, an eval, a failure. I care most about the intersection of software engineering discipline and AI agents, and about judging these systems on evidence instead of vibes.

I also advise engineering teams on building AI systems they can trust. In practice that means evals you can rerun, guardrails that hold up in production, and infrastructure other teams can reuse. Opinions here are my own.