Sub-Agents Don't Make You Faster Until You Do This

I'm building an engineering runtime — humans design and govern, autonomous agents implement. Building it surfaced a reliability problem I couldn't engineer around: if every agent in your pipeline is right 95% of the time and you chain fifty of them, the finished result is correct about 7.7% of the time. Not because any single agent is bad — because reliability multiplies. The exponent is the number of nodes, and it does not care whether you run them in sequence or fan them out in parallel. A fan-out buys wall clock and nothing else. In this video I walk through the whole runtime first — conversation to ADRs to story arcs to PRDs to specs on the human-in-the-loop side, then bounded work packages, a dependency graph, token-budgeted work items, test plans and sub-agents on the automated side. Then I show what happens to all of it without a quality gate, in sequence and in a fan-out. Then the decision that answers it: no worker gets to say it's finished. A deterministic test, written in a separate context from the worker that does the work, makes that call, and the work loops until it passes. Not an LLM judging another LLM — deterministic tests executed in code, because a model asked to check its own work shares the blind spot that produced the work, and will sometimes tell you it's done when it isn't. I also name what it costs, because it isn't free. Every node needs meaningful validators written at plan time, which is human effort before any code exists. Validation costs wall clock on every attempt, including retries — you may wake up to a run that's 70% done rather than nominally finished. Some valuable work has no executable test. And the honest limit, which I haven't solved: the system is only as good as its validators. A weak test suite produces confidently green nodes. This is not a coding tutorial. It's the reliability engineering underneath an agent system, and the decisions I made because of it. CHAPTERS 0:00 The dream, and the reliability truth underneath it 0:36 What I'm building: the human-in-the-loop half 2:55 The fully automated half: work items, the DAG, token budgets 5:47 What happens without a gate: the sequential cascade 7:36 A fan-out looks parallel. It isn't. 9:29 The exponent is the real problem 10:14 50 sub-agents at 95% is a 7.7% chance 11:33 Everyone tests. Is everyone testing correctly? 12:38 The test node, in a separate context 13:47 The quality gate, and waking up 70% done 15:37 How the validators actually get written 17:48 The same graphs, with the gate in place 19:27 What this buys you 20:31 What it costs 21:59 You cannot escape the reliability equation 23:38 What is allowed to be trusted at all The two-hour live build of this system: https://youtube.com/live/jWdjWzA1qs0 AI Software Factory The Engineering Runtime #AIEngineering #SoftwareArchitecture #SystemDesign #AgenticAI #ProductionAI #ReliabilityEngineering