3 Agents, Same URL Error, 1 Gate Stopped All 3

The Pattern That Broke Three Agents

Here's what happened on August 8, 2026. Three different agents (Hermes, C005-capability-probe, and C005-2) all delivered HTML review panels with the exact same wrong port: 8443.

The correct port was 8444. It was written down. It was known. And yet, every single agent chose 8443 anyway.

This is one of those things that feels absurd in hindsight but is incredibly common in practice. Knowledge ≠ execution. An agent can "know" the right answer in its context window and still produce the wrong one because there was no gate between thinking and delivering.

Event #1: E-SHUYU-004 — The First Wake-Up Call

The Output Compliance Checker v2 (OCC v2) auto-detected the wrong port while checking a capability-probe gate file. The check ID was url-b7a2b2a1ba, failure detail: WRONG_PORT.

Here's what made this event significant: the gate caught it before any human saw it. For the first time, LAO took on the trust cost itself — intercepting a delivery error before it reached the founder. This generated an automatic future anchor (future-anchor-1786200319) through OCC v2's self-harvesting loop.

Impact: +0.3. We'd gone from "agent makes mistake → human discovers → human fixes" to "gate detects → gate blocks → system learns."

Event #2: E-SHUYU-005 — Even After Success, Errors Repeat

This was the harder lesson. Four times, HTML review panels were successfully delivered with the correct URL. Everything looked good. Then C005-2 did it again — same pattern as before: claimed the URL existed, file returned 404, fixed it, then got the POST field name wrong too.

The founder's question cut to the core:

"You've submitted the review successfully so many times today. Why isn't the experience getting better?"

That's not criticism. That's the most honest engineering observation possible. Past success doesn't create muscle memory. Without explicit enforcement, humans and agents alike revert to old patterns under pressure.

How OCC v2 Emerged as the Single Gate

By the end of the day, we had documented the full arc across both events:

Individual failure — Hermes used wrong port [V]
Auto-detection + blocking — OCC v2 caught E-SHUYU-004 before delivery [V]
Repetition despite success — C005-2 repeated in E-SHUYU-005 [V]
Escalation to checklist — C-FOUNDER-REVIEW-DELIVERY-CHECKLIST created [V]

The progression wasn't linear. It was messy. But each step hardened the system further.

S-LAO-002: The Real Breakthrough

What made this whole episode matter was happening in parallel behind the scenes. OCC v2's Intent Validation Gate v2 ran five Platform Trap tests (TF001–TF005) and auto-generated:

103 future anchors for output compliance domain (URL ports, Feishu zero-tolerance, output formatting, pre-delivery checks)
11 future anchors for intent validation domain

That's 114 automatic hardening anchors [V]. Each one represents a potential failure mode captured before it could affect a user. The system is literally learning from itself faster than new bugs can be introduced.

The New Checklist

For E-SHUYU-005, we built the C-FOUNDER-REVIEW-DELIVERY-CHECKLIST: a 6-item mandatory pre-delivery verification list:

File actually exists (not just URL claimed)
Port is 8444 (not 8443)
curl returns HTTP 200
POST field names are correct
Interactive elements are present
C005-2 validator passes

Every delivery now goes through this gauntlet. No exceptions. Not because we don't trust our agents — because we know they'll forget. The checklist doesn't replace intelligence; it augments it with discipline.

What This Teaches About Automation

Most teams build automation to reduce work. We're building ours to increase reliability. There's a difference.

Automation that just speeds up a broken process gives you a broken result faster. Automation that catches errors before they reach the user gives you a system that gets more reliable over time.

Our progression this day:

The last row is where this gets interesting. Once an error triggers a future anchor, the probability of the same class of error drops toward zero. The system compiles its own experience into code.

Open Questions

Is a checklist the right solution for E-SHUYU-005? Or should we have built a stronger gate earlier? Probably both. Gates catch structural errors. Checklists catch procedural slips. You need both.

Would you rather have an intelligent agent that occasionally messes up, or a less flexible system that never does? We found out you can have both — if you layer them correctly.

TL;DR: Three agents, one port number, endless frustration. OCC v2 emerged as the single gate catching all instances, generating 114 automatic hardening anchors. Post-success repetition led to a mandatory 6-item delivery checklist. Automation isn't about speed — it's about making systems more reliable with every failure.

Try LAO Yourself

LAO turns agent failures into permanent prevention. Every error gets detected, repaired, anchored — and never repeated. The OCC v2 alone has already generated 114 self-hardening anchors.