Page 24
12 articlesTime-to-First-Token Is the Latency SLO You Aren't Instrumenting
p50 and p99 total latency miss the single number that governs how your AI product feels: time to first token. Here is why reasoning models make it worse, what to measure, and how to route around it.
The Unmergeable Agentic Refactor: Why Multi-File Diffs Break at the Seam
Agent-authored refactors look clean file-by-file and fail at the seams. Why hunk-level review misses cross-file bugs, and the compile-first, program-analysis discipline that fixes it.
The Validator Trap: How Post-Hoc Guards Rot Your Prompt From the Inside
Every output validator you bolt onto an LLM feels like a fix. Over time, those fixes rewrite your prompt into a defensive contract that starves the model of reasoning capacity. Here is how to audit and refactor the damage.
Voice Agents Are Not Chat Agents With a Microphone: The Half-Duplex Tax
Voice agents inherit the human half-duplex protocol, not the comfort of chat. Turn negotiation, barge-in, and the real 200ms budget decide whether your agent sounds attentive or uncanny.
Agent Fleet Concurrency: Coordinating Dozens of Agents Without Deadlock or the Thundering Herd
A fleet of LLM agents is a small distributed system, not thirty copies of one agent. Admission control, AIMD backpressure, circuit breakers, and external-state coordination are what keep a concurrent fleet from eating itself.
The AI Changelog Problem: Why Your Prompt Updates Are Breaking Other Teams
Prompt edits, model upgrades, and tool schema tweaks change behavior without changing code. Here is the changelog format and versioning contract that keeps consuming teams unblocked.
Design Your Agent State Machine Before You Write a Single Prompt
The instinct to start with a prompt and bolt on logic afterward creates agents that work once and fail mysteriously in production. Here's how designing state machines first changes everything.
The Data Rollback Problem: Undoing What Your AI Agent Wrote to Production
Traditional database rollbacks break when AI agents write to production at machine speed across distributed systems. Here's the architectural shift required to make agent-written state recoverable.
The Attribution Gap: How to Trace a User Complaint Back to a Specific Model Decision
When users report wrong AI advice, most teams can't reconstruct which model version, prompt, or retrieved context produced that output. A guide to the logging schema, trace propagation, and sampling strategies that make AI complaints investigable.
The AI Audit Trail Is a Product Feature, Not a Compliance Checkbox
Showing users what your AI agent actually did — which tools it called, what data it retrieved, where it branched — increases adoption more reliably than any feature flag experiment. Here's how to build it.
AI Code Review in Practice: What Automated PR Analysis Actually Catches and Consistently Misses
AI code reviewers catch typos and null checks at 70-85% accuracy but miss semantic errors 85-90% of the time. Here's the empirical breakdown and the workflow design that avoids turning automated approval into a rubber stamp.
AI Compliance Infrastructure for Regulated Industries: What LLM Frameworks Don't Give You
Finance, healthcare, and legal deployments require immutable audit logs, output lineage, refusal tracking, and explainability hooks that most LLM frameworks don't provide out of the box. Here's the architecture that fills the gap.
About Tian Pan
I'm Tian Pan, an engineer-founder focused on agentic engineering — building autonomous AI systems and scaling engineering teams. I write practical guides on system design, technical leadership, and shipping with AI agents. Previously an early engineer at Uber, Brex, and IoTeX.
