Why I Let AI Fight With Itself Before I Ship Anything

JLJeff Liu··2 min read

I build with AI most of the day. Multi-agent systems, content pipelines, infrastructure decisions, database schemas. It's my primary toolchain.

And I ran into a problem.

The longer one AI ran in a single session, the more the output drifted. The answers stayed confident, but the gaps got bigger, and harder to spot because everything still looked clean.

The two-pass discovery

Today's task was a persistent memory system for the platform. Claude Desktop (Sonnet 4) came back with a plan. Clean architecture, three layers, a timeline of 5-8 days.

Then I asked Google Antigravity (Claude Opus 4.6) to review it.

Antigravity came back with a table of gaps. The embedding model was assumed, not decided. The vector index strategy was not addressed. There was no conflict resolution for when two agents write contradictory memories about the same preference.

The compression prompt quality was hand-waved. And the timeline was flagged as a prototype estimate, not a production one. The real number is 2-4 weeks.

Claude Desktop's plan was not wrong. It was incomplete. And confident about being incomplete, which is the part that gets you into trouble.

Different models, different blind spots

This is not a knock on Claude Desktop. I use it every day. But every model has a perspective shaped by how it was trained and what it was optimized for.

Claude Desktop (Sonnet 4) optimized for clarity and structure in a consumer chatbot context. Antigravity (Claude Opus 4.6) pushed back on assumptions using locally trained institutional data: project history, schema decisions, prior failures.

Neither is the full picture on its own.

The process now

I run a two-pass review on any significant decision. One AI generates the plan. A different AI interrogates it. Not to be adversarial. To be thorough.

This connects to something I've been measuring around context windows and token budgets. Every AI brings context biases into a session. The more specialized the model, the more pronounced those biases get. A model trained heavily on code will pattern-match technical problems differently than one trained on reasoning tasks.

The cost of shipping a plan built on a 5-8 day estimate that should be 2-4 weeks is real. I've shipped plans that looked airtight and fell apart in week two. The plan was reviewed by the same system that wrote it.

Running a second model over the first isn't redundancy. It's the cheapest quality gate I've found for catching the gaps that confidence hides.

Keep reading

Building a Commit Guard for AI Agents That an Adversary Can't Slip Past

I gave a coding agent a guard that checks every commit lands in the repo and branch I intend. My 417 tests were all green. Then I hired a second agent to break it, and it walked through two holes the green suite never thought to check.

Read

Graph Engineering Makes Agent Failure Legible, Not Agents Reliable

A video making the rounds argues graph engineering supersedes the loop-based agent pattern. After building multi-agent systems, I think the loop-versus-graph framing asks the wrong question. Topology decides whether you can SEE a failure, not whether one happens. Reliability lives in the deterministic checks at the edges, not in how you wire the agents.

Read

Claude Desktop vs Cowork vs Code: A Beginner's Roadmap to Which One to Use

Claude Desktop, Cowork, and Code look interchangeable. The real difference is what survives when the session ends, and that decides which one to build in.

Read

Remote-Controlling AI Coding Agents From Your Phone: What's Real, What It Costs, and Where It's Going

For fifty years, controlling a computer meant sitting at it. AI coding agents broke that. Here''s how I drive my agents from my phone, what it actually costs, the local-vs-cloud distinction, and why this turns you from an operator into a conductor.

Read

Running Multiple AI Coding Agents in Parallel: A System for Safe, Self-Merging PRs

Running several AI coding agents on the same repo turns into a coordination problem fast. Here's the system I built to keep them isolated, review each other's work, and merge their own pull requests safely.

Read

Why I'm Building AI Education for Kids

I didn't think I'd hear back from Runway. Then this morning, I got accepted into their Builders program. Reflecting on why this mission is personal, and how we can ensure kids don't get left behind in the AI era.

Read

AI Native, Human First

You can outsource thinking and execution to AI. But you cannot outsource understanding. Real learning and proactive groundwork are first-class survival skills in the AI era.

Read

Using AI Without Losing Your Humanity

Everyone has access to the same AI tools now. So what sets you apart is no longer how much content you can generate. It's whether it makes people feel something, and whether you can use AI without losing your humanity.

Read

How I Built a Publishing Stack for AI Search

People are finding things through ChatGPT and Perplexity instead of Google. I built a pipeline that handles the optimization for that automatically. Here's how it works.

Read

Building Systems, Then 10x the Output

At Zeta Global, my team gave me room to figure things out. I tried to do the same for the people who came after me, from responsive templates to training engineers across time zones.

Read

45 Rule Files Were Making My AI Worse

I had 45 rule files telling my AI how to behave. The output was getting worse, not better. Then I measured. 170KB of context loaded per session. Here's what happened when I deleted half of them.

Read

How I Turned My AI's Mistakes Into Guardrails

I added Husky because of an anti-pattern. CI because of an anti-pattern. PR requirements because of an anti-pattern. None of it came from reading best practices. It came from breaking things and writing it down.

Read