Akshay Pachaar posts agent production checklist
- Akshay Pachaar shared a production checklist for AI agents on May 23, saying reliable deployment depends on routing, evals, guardrails, tracing and cost controls. - The checklist’s clearest warning was that most failures come from infrastructure “plumbing,” not model logic, with human-in-the-loop patterns included alongside monitoring. - The post was amplified by Avi Chawla on X; the thread remains available on Pachaar’s and Chawla’s social feeds.
Akshay Pachaar’s latest post is useful because it describes AI agents as an operations problem before a model problem. His checklist, shared on May 23 and amplified by Avi Chawla, runs through the components needed to move an agent from demo to production: intent routing, evaluations, guardrails, cost attribution, full-chain tracing, monitoring and human review paths. The framing matches Pachaar’s broader public work. His YouTube bio describes him as a co-founder of Daily Dose of Data Science and a creator focused on LLMs, AI agents and RAG, while his GitHub profile identifies him as a former Lightning AI engineer and developer advocate. ### Why does this checklist focus on “plumbing” instead of the model? Pachaar’s central point, as described in the social briefing and reflected in the checklist themes, is that production blockers usually come from infrastructure rather than the core model. (youtube.com) That means failures in routing, retries, permissions, logs, budgets and handoffs can matter more than whether a team picked one frontier model over another. That emphasis lines up with other recent engineering guidance. Arthur AI’s production-agent checklist says tracing should cover every LLM call, tool invocation, retrieval and key decision point, and it places continuous evaluation and guardrails alongside prompt management and governance. ### What does “intent routing” change in practice? Intent routing is the part of the system that decides what kind of request has arrived and what workflow should handle it. (x.com) In production, that can mean separating simple Q&A from tool-using tasks, escalation cases or requests that need a human approver. Pachaar included routing near the top of the checklist, which suggests he treats orchestration as a first-order reliability issue rather than a later optimization. (arthur.ai) Recent production playbooks make the same case. Future AGI’s guide says agentic applications need explicit step graphs, retries, handoffs and budgets because an agent can traverse many possible paths at runtime, unlike a conventional service with a small number of fixed code paths. ### Why are evals and tracing paired together? Evaluations tell a team whether an agent is performing well; tracing shows why it failed. (x.com) Pachaar’s checklist groups them with monitoring and production readiness, which reflects a common engineering pattern: teams need observable runs before they can build useful regression tests or compare prompt and workflow changes over time. Arthur AI says tracing is the foundation for later experiments because teams cannot isolate whether a bad output came from retrieval, a prompt, a tool call or application logic without end-to-end visibility. (futureagi.com) ### Why include cost attribution in an agent checklist? Cost attribution matters because agent systems spend money across several layers at once: model tokens, tool calls, external APIs and compute. (x.com) Pachaar’s inclusion of cost attribution suggests he sees budget visibility as part of core product readiness, not just a finance afterthought. Prefactor defines agent cost attribution as tracking AI spend back to the agent, team, user or task that incurred it, including tokens, API calls, tool invocations and compute. (arthur.ai) That kind of accounting becomes more important when agents take variable paths and invoke multiple services per task. ### Where does the human fit if the system is supposed to be autonomous? Human-in-the-loop patterns are the backstop in Pachaar’s checklist. (x.com) In practice, that usually means a person reviews high-risk actions, approves external side effects, or handles low-confidence cases that routing or guardrails flag for escalation. MIT-related discussion in the social briefing also pointed to safety, sandboxing and permission controls as essential for coding agents in production, reinforcing the same pattern: autonomy is bounded by review and operational controls. (prefactor.tech) Pachaar continues to publish agent and AI-engineering material through Daily Dose of Data Science and his social channels, where the May 23 checklist post is still available for engineers comparing their own production stack against the items he listed. (x.com) (dailydoseofds.com)
