The Chanl Blog
Insights on building, connecting, and monitoring AI agents for customer experience — from the teams shipping them.
All Articles
235 articles · Page 4 of 20
Circuit Breakers for AI Agents: Stop the 3 AM Meltdown
One retry loop at 11 PM becomes $437 by 7 AM. Here's how to implement circuit breakers for AI agent tool calls, LLM calls, and external APIs, with TypeScript patterns that stop cascading failures before they start.
Past 50 tools, function-calling accuracy falls off a cliff
Past 50 tools, function-calling accuracy falls off a cliff. Measure the curve on your own agent and recover accuracy with per-turn toolset scoping.
GPT-5, Claude 4.5, Gemini Score the Same Calls. Their Kappa Is 0.52
Run the same calls through GPT-5, Claude 4.5, and Gemini and Cohen's kappa lands at 0.52. Here is how to measure judge agreement on your own corpus.
1M-Token Context or RAG? How to Pick for Your CX Agent
Gemini's 1M-token window is real but not free. A practical decision framework for choosing between long-context and RAG for customer experience agents, with cost numbers, code, and the hybrid pattern most production teams land on.
MCP tool description drift: the silent failure nobody alerts on
Edit an MCP tool description for clarity, lose 8% routing accuracy, and the eval suite stays green. How to detect, gate, and roll back the drift.
Your voice agent's P95 is lying. The real problem is P99.9
Per-stage P95 hides the tail customers feel. How variance compounds across STT, LLM, and TTS, and how to SLO the joint distribution.
How to Eval Agents When There's No Right Answer
Most eval methods assume you know the correct response. CX agents rarely have one. Here's how to score agent quality with criteria-based rubrics and LLM-as-judge, no labeled ground truth required.
Stop Loading All Your MCP Tools at Once
Loading 50 MCP tools burns 72K tokens before your agent says a word. Progressive tool discovery fixes that: smaller context, sharper decisions, real code patterns.
How to Build a Voice BDR That Qualifies (BANT as a State Machine)
Most AI SDRs ask 'tell me about your business' and dump benefits. Here's how to encode BANT as tool calls, score live, write to Salesforce, and exit politely.
Build a Save-Desk Voice Agent That Won't Get You Sued
FTC click-to-cancel was vacated. State laws still bite. The cancel-first architecture, one-shot offers, and audit trail for a save-desk voice agent.
How to Build a Cart Recovery Agent (and When to Send Nothing)
Most cart-recovery flows are three discount emails. A real recovery agent decides why the customer left, picks the right channel, caps the discount to protect margin, and sometimes sends nothing at all. Here is how to build it.
Build a Returns Voice Agent That Can't Refund Itself Broke
Returns are 60% of peak ecommerce contact volume. Most voice agents will refund $4,000 on a prompt injection. Here's how to build one that physically can't.
Learn Agentic AI
Weekly. Patterns for shipping agents that work — MCP, scorecards, regression tests, prompts, model comparisons.
