# Agents in production

A pack of 8 sessions from the MLOps Community YouTube channel, in the order to watch them. 4h 01m of video.
Page: https://mlopstalks.com/packs/agents-in-production

An agent calls the right API but changes the wrong record. A restarted process forgets completed work. A run reaches a success page even though the requested task never happened. Start with deployment accounts that expose missing context and unnecessary actions, then narrow the work assigned to the model. Design the tools around explicit arguments, permissions and useful feedback before giving the agent more freedom to plan or execute code. The later talks address crashes, repeated calls and failures that ordinary service monitoring cannot see. Finish with the distinction between reaching the end of a workflow and completing it correctly. Across these 2024 to 2026 accounts, stopping, recovering and asking for help are part of the operating design.

## This pack is for you if

- Your agent works in demonstrations but takes unnecessary or incorrect actions on real tasks.
- Long-running sessions lose progress when a process or external service fails.
- The agent announces success before you can verify the requested change actually happened.

## The talks, in order

### 1. The Agent Landscape - Lessons Learned Putting Agents Into Production

Paul van der Boor & Floris Fok, Prosus Group | 1:08:41 | Agents in Production Series
Video: https://www.youtube.com/watch?v=lRGldru7ohU
Summary: https://mlopstalks.com/talks/the-agent-landscape-lessons-learned-putting-agents-into-production.md

Why first: Real workflows expose assumptions that a controlled demonstration hides: ticket shorthand may depend on unwritten team knowledge, and an extra agent can add a step nobody wants. These deployment accounts establish what must work outside the model before you refine the action loop. Start by making the task and its missing context explicit.

### 2. Linear's Approach to Production AI Agents

Kristin Boyer, LINEAR | 15:15 | MLOps Community
Video: https://www.youtube.com/watch?v=Xq3PrK7V9Vw
Summary: https://mlopstalks.com/talks/linears-approach-to-production-ai-agents.md

Why second: Boyer gives that discipline a concrete implementation. Early routing narrows the available actions, nearby conversation supplies context, and fixed business rules move into application code. The model gets the language work it can help with. That smaller responsibility makes it possible to design precise tools instead of explaining every exception in a prompt.

### 3. Building Agentic Tools for Production

Sam Partee, Arcade AI | 23:55 | Agents in Production 2025
Video: https://www.youtube.com/watch?v=8TRQo3mvl7g
Summary: https://mlopstalks.com/talks/building-agentic-tools-for-production.md

Why here: Partee separates read-only exploration from narrowly defined actions and gives each tool its own authorization requirements. Constrained arguments and useful error responses make the model's next step easier to choose. With the task already narrowed, this turns a general API connection into an interface whose effects you can inspect.

### 4. Building Reliable Agents

Eno Reyes, Factory.ai | 24:45 | MLOps Community
Video: https://www.youtube.com/watch?v=4jCkcNOQPmU
Summary: https://mlopstalks.com/talks/building-reliable-agents.md

Why here: Reyes connects tool design to planning and environmental feedback. Explicit success criteria keep the goal visible, while processed logs give the agent evidence it can use without flooding its context. This explains what a reliable loop needs before the next talk considers giving that loop a more powerful action format.

### 5. Deploying Executable Agent Workflows

Gal Peretz, Carbyne | 18:50 | MLOps Community
Video: https://www.youtube.com/watch?v=2N-mnKLE0Lk
Summary: https://mlopstalks.com/talks/deploying-executable-agent-workflows.md

Why here: Peretz compares individual tool requests with generated programs that express loops and aggregation directly. The extra freedom can help large branching tasks, but it requires a restricted execution environment and inspection before code runs. Treat this as a specific option when fixed calls become awkward, not an automatic upgrade for every agent.

### 6. Building AI Agents That Survive Production

Haytham Abuelfutuh, Union.ai | 31:59 | AI Agents 2026
Video: https://www.youtube.com/watch?v=swO5svhBhQ4
Summary: https://mlopstalks.com/talks/building-ai-agents-that-survive-production.md

Why here: Abuelfutuh shifts from deciding what to execute to preserving it when infrastructure fails. Recorded actions and durable session state let his proposed system recover without repeating earlier model calls or user questions. This is a distinct failure from bad reasoning: a sound action loop still needs a way to survive a disappearing process.

### 7. Building Replit Agent - Hard Lessons Learned

 | 32:31 | MLOps Community
Video: https://www.youtube.com/watch?v=RYde73eO7ok
Summary: https://mlopstalks.com/talks/building-replit-agent-hard-lessons-learned.md

Why here: A running process can still contain a failing agent. This coding-agent account uses rollback behavior, user frustration and traces to find loops and unwanted detours that ordinary health monitoring misses. After making execution recoverable, add evidence that tells you whether the ongoing work is useful enough to continue.

### 8. Catastrophic agent failure and how to avoid it

Edward Upton, Asteroid | 25:21 | Agents in Production 2025
Video: https://www.youtube.com/watch?v=wsZdGllOAGc
Summary: https://mlopstalks.com/talks/catastrophic-agent-failure-and-how-to-avoid-it.md

Why last: Upton's appointment-booking failure shows why a generic success page is insufficient evidence of completion. Scoped workflows, visible human review and an independent inspection of the execution provide ways to catch that mistake. End with a controlled stop or escalation when the alternative is an unjustified success claim with real consequences.