Technology
Agent Infrastructure, Silent Failures and the Open Source Gap
Ran Aroussi explains why AI agents should run on server-style infrastructure, where agent systems fail without warning, and why open source licensing has not caught up with the cloud.
E01: AI and the Open Source Frontier | Old School / New Tech
Video from VarOps.
Generated from the canonical interview transcript and validated against source data by Guests on Air.
How should teams build and license production AI agent infrastructure?
Treat agents like any other workload that runs on a server rather than rebuilding them inside a new framework for every project. Instrument them heavily so every decision and tool call can be traced, add a fact-checking step to catch hallucinations, and choose a license that keeps the code open while stopping cloud giants from reselling it as a hosted service.
Ran Aroussi has spent the past year heads down, building agent infrastructure, writing a book and running a software agency. On the relaunch of the Old School New Tech podcast, recorded live with an AI co-host called Maximus, Aroussi explained the thinking behind Muxy, an open source server for AI agents, and what it has taught about running agents in production.
The conversation covered three connected ideas. Agents should run on shared infrastructure, the way websites run on a web server, instead of being rebuilt inside a fresh framework each time. Production agent systems fail in predictable places, and teams need to see inside them to catch it. And the licenses that govern open source code were written before hyperscalers existed.
For founders and engineering leaders deciding how to put agents into real workflows, the episode works as a practical checklist. Pick infrastructure that behaves like a server, instrument it until every decision can be traced, add a verification step for anything a model produces, and think hard about the license before releasing code that a business depends on.
Key takeaways
- Run AI agents on shared server-style infrastructure instead of rebuilding them inside a new framework for every project.
- Observability, traceability and debuggability are where agent systems fail quietly, so trace every decision and tool call.
- Hallucinations need a process fix: add a final step that verifies and fact-checks data before any answer goes back.
- Permissive licenses let hyperscalers resell open code as a service; a SaaS-protective license keeps code open without that risk.
1. Build agents like servers, not frameworks
Aroussi built Muxy because the agent ecosystem felt crowded with frameworks that missed the point. Teams were writing the same plumbing again on every project, wiring memory, tools and orchestration by hand. Coming from a systems background, Aroussi saw a familiar pattern. Earlier generations of software solved the same problem long ago by moving common work into dedicated servers that applications simply plug into.
The comparison is the web server. Nobody building a website writes an entire web server first; they use Nginx, Caddy or Apache and spend their effort on the application itself. Gaming servers and rendering engines follow the same logic. Aroussi wants agents treated the same way, with the hard primitives living in one dependable layer, so an application hooks into a service rather than rewiring everything from scratch.
That is the idea behind the podcast's name, Old School New Tech: bring the discipline of older infrastructure to agentic AI. For teams, the practical gain is reuse. A shared runtime means fewer bespoke builds, fewer places for things to break, and a cleaner line between the agent work that is specific to a business and the plumbing that every agent needs anyway.
2. Find the places where agent systems fail silently
Asked where this kind of structure fails without anyone noticing, Aroussi named the weak spots. Observability and traceability come first, and they are closely related. Next is debuggability, the ability to follow an agent's chain of thought and see why it did what it did, which tools it called, and what led it to choose those tools over the others available.
Muxy's answer is a large volume of signal. Aroussi said the platform now emits close to 400 observability events, enough to trace everything happening inside a request. Today those events print to standard output or flow into tools such as Datadog, Splunk or ClickHouse. The missing piece, which Aroussi plans to ship soon, is a proper interface for digging into full traces without stitching logs together by hand.
The last weak spot sits outside any runtime: model hallucinations. No server can stop a model inventing facts, so Aroussi handles it with standard operating procedures. The final step of a workflow should verify the data and fact-check the answer before it comes back to the person who asked. That is a process control rather than a technical trick, and it belongs in every production agent design.
3. Rethink open source licensing for the cloud era
The sharpest part of the episode was about licensing. Aroussi released around 30 repositories under Muxy, all under permissive Apache or MIT licenses, but chose the Elastic License V2 for the main runtime. That license lets anyone use, modify and redistribute the software, including for commercial purposes, but blocks offering it as a managed service. The target is hyperscalers such as AWS and Google Cloud.
Aroussi argued that letting a vendor take a project, fork it, keep the improvements private and sell it as a hosted service does no justice to the open source world. The original spirit, going back to the Homebrew days of the late 1970s and early 1980s, was sharing knowledge rather than cannibalising someone else's work. Aroussi was clear that the license choice is not a money grab.
Aroussi wants the Open Source Initiative to bless a new license close to MIT, fully permissive except for offering the software as a service, similar to the license that DHH and his company adopted for a recent product. Sentry's fair source effort was, in Aroussi's view, the closest honest attempt, but it did not catch on. The question the episode left open is whether the initiative can embrace that model without losing its core values.
