As of July 2026, four separate agent tools landed on Hacker News inside the same 48-hour window, and none of them were pitching a smarter model. They were pitching proof that your agent is already failing you, quietly, with no error message, no downtime alert, and no dip on a status page. Agnost AI, a Y Combinator S26 company, reads through agent conversations looking for what its founders call rageprompting, customers cursing at the bot, plus repeated rephrasing and mid-conversation corrections. Oodle.ai launched full agent trace storage priced at $10 per million traces, arguing that sampling hides exactly the failures that matter. FixBugs built an agent whose only job is reproducing a bug in a sandbox and verifying the fix before it ships, because a patch that looks right in a code review is not the same as a patch that works on the next real customer. Rejourney, built by a UT Austin sophomore, takes the same logic to session recordings, predicting where a web or mobile app is quietly leaking revenue before a customer ever files a complaint. If your business runs a booking assistant, a voice receptionist, or an outreach agent, this is the pattern worth learning from: uptime is not the same as working, and the gap between the two is where customers quietly leave.
Why this matters for a local business
An agent does not have to crash to fail a customer. It can answer instantly, sound coherent, and still quote the wrong price, skip the booking step, or loop a caller through the same question three times until they hang up and call the next business on the list. None of that trips a server alert, because the server is doing exactly what it was told to do. The only record of the failure is in the conversation itself, and most businesses never read their own transcripts. Oodle's $10-per-million-traces price point matters here because it removes the cost excuse that used to be legitimate. Full logging of a small business's agent volume, thousands of conversations a month for a single location, runs a few dollars. The businesses still flying blind on this are choosing to, not budgeting around it.
The risk is highest on the agents doing the most talking: a voice receptionist fielding after-hours calls, a booking agent handling scheduling back-and-forth, and an outreach agent following up on leads by text. Each of these has dozens of small decision points per conversation, and each decision point is a place where the agent can quietly go off-script while still sounding fine on the surface. A chat widget that only answers FAQs has far less room to fail than an agent negotiating a time slot or a price.
The playbook: five steps to catch a silently failing agent
- Log every conversation in full, not a sample. Sampling is a habit carried over from expensive observability tooling built for a different era. At current pricing, storing full transcripts for a single-location business costs less than a week's worth of coffee for the front desk. Sample logging means the one conversation where a customer got quoted double your rate is exactly the one you never see, because statistically it was never going to land in the sample.
- Search transcripts for the four behavioral red flags. Agnost's approach is worth copying even without their software: scan for cursing or frustrated language, a customer repeating the same request in different words, a customer correcting the agent mid-conversation, and any request to speak with a human. Each of these is a customer telling you, in real time, that the agent did not understand them the first time, and most of them will not tell you a second time. They will just leave.
- Set a same-day escalation rule, not a monthly review. A conversation with two rephrasings or one correction should route to a human within hours, not sit in a queue until someone has time to look. A booking that fell through on a Tuesday afternoon is a customer who already booked with a competitor by Thursday morning.
- Reproduce before you patch. FixBugs' model applies directly to an agent stack: when a conversation fails, replay the exact customer inputs in a test environment before touching the live prompt or workflow. Patching blind off a single support complaint is how one fix quietly breaks three other conversation paths that were working fine yesterday.
- Tag every flagged conversation with what it was worth. Not every failure costs you the same amount, and this is where the Rejourney approach to revenue-leak prediction is worth borrowing: rank failures by dollars at risk, not by ticket volume. A rephrased question that still ended in a booked appointment is noise. A correction that ended in a hang-up on a job worth $2,000 is the one to fix first, this week, not whenever the backlog clears.
Quick self-audit checklist
Run through this before you assume your agent stack is fine:
- Do you log 100% of agent conversations, or only a sample?
- Can you search transcripts for phrases like "let me talk to a person" or "that's not what I asked"?
- Does a flagged conversation get reviewed the same day, or does it wait in a queue?
- When you fix a failure, do you replay the original conversation to confirm the fix, or ship and hope?
- Do you know which flagged conversations cost you a booked job versus which were just a customer being chatty?
Common pitfalls
- Treating uptime as the success metric. An agent that answers every call and books nothing is technically "up" and functionally useless. Uptime tells you the lights are on; it tells you nothing about whether the conversation actually worked for the customer on the other end.
- Sampling to save money. At $10 per million traces as a market benchmark, the businesses still sampling are optimizing for the wrong line item. The failure you don't log is the failure you never fix, no matter how good your team is at responding to the ones you do catch.
- Shipping a prompt fix without reproducing the original failure. A fix that isn't tested against the exact conversation that broke it can pass a quick review and still fail the next customer who phrases the same request slightly differently.
- No one owns the review. Tooling that flags failures does nothing if no person is assigned to read the flagged transcripts every week. This is the single most common reason monitoring gets set up once, looks good in a demo, and is quietly ignored three months later.
This is the same blind spot we've written about before: an agent stack can fail without ever showing up as downtime, and the businesses catching it early are the ones actually reading their own transcripts instead of trusting the dashboard to tell them everything is fine.
Before you can fix what your agents get wrong on a call or a chat, it helps to know how your business shows up in the first place when a customer asks ChatGPT, Claude, or Perplexity for the best option near them. Get a free AI Visibility Report and see exactly what those tools say about your business today: free AI Visibility Report.