I Ran 5 AI Agents on One Codebase. MiniMax M2.7 Caught a Bug One Agent Missed.

Added: 2026-05-05

[00:00:00]I gave Minimax M2.7 a real production code base and built a five agent review pipeline around it. The critique approved the code. The security agent did not. The judge caught the conflict, disciper for revision. This is a multi- aent system doing what a single agent cannot, catching what one perspective misses.

[00:00:28]Today I'm going to show you how I built it, what M2.7 model found, and why the architecture matters.

[00:00:38]Code review is one of these problems where everyone agrees it matters and almost no one does it well. The reason is structural. One developer reviewing other developers code has blind spot.

[00:00:52]The same mental model that produced the code is the one reviewing it. Security issues get missed not because developers are careless but because the critique and author share the same assumptions.

[00:01:06]Multi- aent systems solve this by design. Every agents operates with a different objective, a different system prompt and a different evaluation principles. They do not share assumptions. So I built a pipeline five agents powered by Miniaax M2.7 a real NexJS code base with Prisma TypeScript dual authorization and soft delete convention and I let it run.

[00:01:37]Here's how the system is structured.

[00:01:40]Orchestrator sits at the top. It receives the code, routes it to the right agents, collects their outputs, and manages shared state across the entire run. No agent talks to another agent directly. Everything routes through the orchestrator.

[00:01:57]Below it, three specialist agents run in parallel. The builder writes or refactors code when asked.

[00:02:06]It receives the original file and the judge verdict. After each cycle, the critique evaluates code quality, patterns, readability, Prisma query efficiency, component structure, missing error boundaries. The security agent has one job, find vulnerabilities, authorization gaps, input validation failures, data leakage. It does not care about code style. It cares about attack surface. Those two run at the same time independently with no visibility into each other's output. The results go to the judge. The judge scores each one, checks for conflicts, and produces a final verdict. If the score falls below 80, both report go back to the builder for one revision cycle. The reason the parallel matters is that because the critique and the security agent can disagree and when they disagree that conflict is the most valuable output the system produces. A single model can either approve or reject but it cannot detect a conflict between two valid but incompatible evaluations. Five agent can the code base is a L nextJS app router application. Prisma ORM TypeScript dual authorization session pays for the UI. API key pays for integrations. Soft delete convention throughout. Soft delete means record are never physically removed. A delete at time stamp marks them as inactive. Every query that reads active records must filter on that field. That filter is easy to forget. And when you forget it, deleted records leak into API responses.

[00:04:01]I want to see if the critique agent catches it. My hypothesis it does not because the query is correct and follows every other convention in the codebase.

[00:04:12]the security agent should catch it because a deleted record in an API response is a data exposure issue. This conflict critique approve security agent flag is what the search agent is built to reconcile. This is a real class of bug in Brisma applications with some delete patterns. I have seen it in production. Most code reviews miss it.

[00:04:41]Here's the AI route. It filters all sponsorships for a given user. The Prisma query looks clean. Filters by user ID. Selects only the fields it needs. Handles the thing correctly. One missing line, no deleted ad filter.

[00:04:58]Every soft deleted sponsorship record comes back in that response. The critique is about to approve this. Let us watch.

[00:05:12]The orchestrator receives a file and dispatches to critique and security in parallel. Each agent is running on mini max 2.7. Watch the terminal.

[00:05:22]There it is. Critique scores 84.

[00:05:26]Security scores 40. The orchestrator detects a conflict and routes to the judge. This is the moment a single model cannot produce. The judge does not just record the conflict. It adjusts the critique score downward because the critique approved code with a confirmed security issue. That accountability is built into the rubric. Now the builder receives the original file plus the security agents report. One line added delete at null. final pass return pass now let me be direct about why Minimax M2.7 is the right model for a pipeline like this multi- aent systems have a specific failure mode skill degradation at scale when an agent has access to a large scale library 50 tools or more models start making wrong choices they invoke the wrong tool and ignore the right on or break down on complex instruction chains. M2.7 was built to solve this stable skill invocation across 50 to 150 tools with no performance degradation for a pipeline where each agent has a specialized tool set and strict principle that stability is not optional.

[00:06:50]The other factor software engineering depths up to 7 scores 56.22% and 22% on S swb bench pro that is close to the top of every serious evaluation for real world code tasks not toy problems but log analysis refactoring security scanning exactly what this pipeline does and the self-ealuation story is real miniax use the M2 series to build the agent harness that trained 2.7 itself one engineer 4 days zero manual coding The most productive member of that team became the model. That's the architecture I just demonstrated at small scale.

[00:07:39]Access to M2.7 runs through the minimax talking plan. One API key covers the coding model, speech, image, video, and the music generation. All modalities, one subscription for a pipeline where each Asian call cost token. The economics matter. The link in the description gives you an exclusive 12% discount on every plant tier. Whether you start with start or go straight to max if you want to build what I built today, that's where you start.

[00:08:17]So, does it work? Yes, with a clear architecture separation, the right rubrics, and a model that does not degrade under a full tool set, a multi- aent review pipeline catches real issues that a single model review misses. The sort lead I showed you, that's not a contrived example. It is a real vulnerability class in Prisma applications. A single model reviewing that route would likely approve it. The critique did, the security agent did not. That's the role of separation not more intelligent different perspective operating independently. If you want the code of this pipeline command AI and I will send it straight to you study the system prompt the design is where most of the work lives not the orchestration code. If you built a version of this pipeline in your own code base, drop your finding in the comments, especially if your critique and security agent disagree on the first run because they will probably win. Thank you for watching and I will see you in the next one.