Day 2 · Recorded 7 May 2026
Dark Factories: verification-first software delivery with agent fleets
If background agents create more code than humans can read, the next platform problem is proof. Shardul Vaidya argues that useful agentic engineering depends on dark-factory architecture: specialized agents, isolated sandboxes, DAG planning, adversarial review, deterministic gates, and enough back pressure that the system can force agents to fix what they got wrong.
Shardul Vaidya, Solutions Architect, Data and AI, AWS
What's in this session
Dark Factories borrow a manufacturing idea: once humans are no longer required inside every production step, the "lights" can go out. For software, Shardul's claim is not that judgment disappears; it moves into requirements, orchestration, tool curation, and verification systems.
In this session, he maps the converging architecture: requirements come from Slack, Linear, GitHub, or a web UI; an orchestrator decomposes them into a DAG; agents run in isolated sandboxes with curated tools; plan and code adversaries apply back pressure; deterministic checks decide whether work moves forward or loops back into rework.
For engineering and platform leaders, the takeaway is a blunt one. Monolithic "trust all tools" agents are not the path to scale. The durable advantage is a factory that can plan, dispatch, verify, and learn from failure while keeping humans focused on system design and correctness.
Inside the recording
- 00:00 What a dark factory means for software
The goal shifts from making humans faster to building systems that can verify agent work. - 04:00 The convergent dark factory architecture
Sandboxes, curated tools, sub-agent orchestration, and verification gates show up across teams. - 08:00 DAG planning and rework loops
Failed work is fed forward as context, turning retry into structured rework. - 10:00 A live factory from requirements to tasks
Shardul shows a dashboard that turns a requirement into planned, parallelizable work. - 14:00 Plan adversaries, code adversaries, and back pressure
Review agents and deterministic tests force the factory to listen before anything merges. - 16:00 From writing code to composing the factory
The engineer's job becomes requirements, tool curation, verification, and correctness.
More sessions on agent infrastructure
- Spectre: Harvey's collaborative cloud agent platform — Joey Wang, Harvey
- Backgrounding the Toil: Inside Uber's Agent-Ready Developer Platform — Nikhil Ramakrishnan, Uber
- Enabling AI Tools Without Losing Control: Lessons from a Regulated Bank — Suhail Patel, Monzo