Uploaded October 2026 | Updated November 2026, 3 hours ago
Running autonomous AI coding agents directly on your local machine or shared development server is a major security liability. Unmonitored LLM agents executing arbitrary terminal commands can accidentally delete local files, leak SSH keys, or corrupt repositories. Furthermore, running agents in terminal UIs ties execution directly to your laptop, meaning work stops the moment you shut the lid.
In this technical breakdown and live demo, James (Community Ambassador at Daytona) demonstrates how Daytona AI Code Sandboxes solve these security and scalability bottlenecks. Learn how to run background coding agents inside ephemeral cloud containers, use JSON streaming mode for programmatic agent control, isolate parallel agent branches in Git, and schedule daily automated code hygiene PRs.
Chapters:
00:00 Introduction to Daytona AI Sandboxes
01:05 Understanding Background Agents
02:48 Daytona's Core Problem & Solution
05:01 Visualizing a Daytona Sandbox
05:51 Building Background Agents with SDK
09:51 Live Demo: Collaborative AI Agents
17:04 Scheduled Agents & Code Hygiene
18:38 Try Daytona & Community
Resources:
🔬 Phoenix (open source): phoenix.arize.com
đź”— Arize AX: arize.com
đź“– OpenInference: github.com/Arize-ai/openinference
đź“– Phoenix docs: docs.arize.com/phoenix
How is your engineering team isolating autonomous coding agents in production? Share your setup in the comments below!
Subscribe and hit the notification bell to track modern AI developer infrastructure with Arize AI! đź”” youtube.com/@arizeai?sub_confirmation=1
#AIEngineering #AICoding #DevContainers
Running autonomous AI coding agents directly on your local machine or shared development server is a major security liability. Unmonitored LLM agents executing arbitrary terminal commands can accidentally delete local files, leak SSH keys, or corrupt repositories. Furthermore, running agents in terminal UIs ties execution directly to your laptop, meaning work stops the moment you shut the lid.
In this technical breakdown and live demo, James (Community Ambassador at Daytona) demonstrates how Daytona AI Code Sandboxes solve these security and scalability bottlenecks. Learn how to run background coding agents inside ephemeral cloud containers, use JSON streaming mode for programmatic agent control, isolate parallel agent branches in Git, and schedule daily automated code hygiene PRs.
Chapters:
00:00 Introduction to Daytona AI Sandboxes
01:05 Understanding Background Agents
02:48 Daytona's Core Problem & Solution
05:01 Visualizing a Daytona Sandbox
05:51 Building Background Agents with SDK
09:51 Live Demo: Collaborative AI Agents
17:04 Scheduled Agents & Code Hygiene
18:38 Try Daytona & Community
Resources:
🔬 Phoenix (open source): phoenix.arize.com
đź”— Arize AX: arize.com
đź“– OpenInference: github.com/Arize-ai/openinference
đź“– Phoenix docs: docs.arize.com/phoenix
How is your engineering team isolating autonomous coding agents in production? Share your setup in the comments below!
Subscribe and hit the notification bell to track modern AI developer infrastructure with Arize AI! đź”” youtube.com/@arizeai?sub_confirmation=1
#AIEngineering #AICoding #DevContainers









![How Uber Evaluates AI Agents at Production Scale | Arize Observe 2026
Running evals is one thing. Building an evaluation system that uncovers problems, changes the roadmap, and continuously improves an AI agent is much harder.
In this Arize:Observe 2026 session, Aayush Agrawal, Senior AI Product Manager at Uber, explains what Uber learned while building an evaluation platform for production AI agents.
Uber’s agent platform supports teams ranging from first-time agent builders to engineers shipping customer-facing agents at global scale. Across those teams, Uber repeatedly found that access to evaluation tools was not enough. Teams needed tracing by default, automatically generated evaluators, continuously updated datasets, broader ownership, and a development process built around learning from production.
Aayush covers:
• Why teams tend to build their agents first and add evals later
• How Uber provides tracing automatically when an agent is deployed
• Why complete agent trajectories matter more than input-output logs
• How Uber generates agent-specific evaluators from configurations and traces
• Turning production failures into continuously updated evaluation datasets
• Using real conversations to simulate multi-turn agent behavior
• Giving product managers, designers, and operations teams ownership of evals
• Why optimizing for a single launch score can produce misleading results
• The questions Uber uses to assess whether an evaluation system is useful
• How a voice-booking agent exposed a failure that offline evals missed
• Using changes in session length to detect unexpected production behavior
• How production traces can power agent insights, experiments, and improved versions
One example came from Uber’s voice-booking agent. When a child mentioned wanting pizza during a ride request, the agent interpreted the background speech as a new destination. Offline evals had not anticipated the scenario, but a spike in the average number of conversation turns surfaced the problem. A conversational designer then used that insight to update the agent, evaluators, and dataset.
The larger lesson is that evals should help teams learn what to build next. The most effective systems connect production traces, failure analysis, datasets, experiments, and agent improvements in one continuous loop. :contentReference[oaicite:0]{index=0}
Chapters:
00:00 Why having evals is not enough
00:48 The stakes of running AI agents at Uber
01:40 Inside Uber’s agent platform
02:41 Supporting every type of agent builder
03:38 Why teams add evals too late
04:25 Making tracing the default
05:48 Automatically generating useful evaluators
07:03 Building datasets from production failures
08:20 Bringing product and design teams into evals
09:37 Why launch-gate metrics fail
10:09 Better questions for evaluating your evals
12:01 What Uber’s voice-booking agent taught the team
13:12 From eval afterthought to product insight engine
13:36 The future agent improvement loop
15:01 The hardest part was never the tooling
đź”— Learn more about Arize: https://arize.com
đź”— Explore Arize:Observe: https://arize.com/observe
đź”” Subscribe for more talks about AI agents, LLM evaluation, observability, and production AI:
https://www.youtube.com/@arizeai?sub_confirmation=1
#AIAgentEvals #Uber #AIEngineering How Uber Evaluates AI Agents at Production Scale | Arize Observe 2026](https://i.ytimg.com/vi/vJh126DQzEc/mqdefault.jpg)
