← All themes and ideas
Agent reliability confidence: medium score 6.9/10 2026-11-06

Coding agents intermittently wander and stall; no way to tell when to trust which

The need

A user reports Codex has been acting 'spaced-out': looping, saying it will do something then not doing it — and is falling back to Gemini/Claude case by case with no signal of when each agent is reliable.

Original post — verbatim Thai

ทำงานแบบเอ๋อๆ พาวน บางทียอกจะทำแล้วก็ไม่ทำ กลับมาซบอกพี่geminiกับclaude

Group member · 2026-11-06  ·  open on Facebook ↗
Same pain, other posts

อยากเห็นเคสที่พังจริง ๆ ว่าแต่ละคนเจออะไรกันมาบ้างครับ

Group member · 2026-11-14 · original post ↗
Three business opportunities
OPPORTUNITY 01

Agent health monitor

What
A CLI wrapper that logs runs across coding agents, scores completion rate and detects loops/stall patterns, publishing 'which agent is solid today' guidance.
Who pays
Professional devs and small studios pay a subscription for routing confidence.
Price
฿149–399/mo — hypothesis
Why now
Multi-agent fatigue is visible across this group's posts; nobody aggregates reliability telemetry.
Effort
M
Moat
A longitudinal cross-agent eval dataset that compounds.
Risk
Users may tolerate free heuristics; per-agent telemetry is fragile across releases.
OPPORTUNITY 02

Auto-routing agent gateway

What
Proxy that dispatches coding tasks to Codex/Claude/Gemini by task type with retry, loop detection and guardrails.
Who pays
Teams pay per seat for higher throughput and fewer dead loops.
Price
฿99–299/seat/mo — hypothesis
Why now
The poster's manual model-hopping is exactly what routing automates.
Effort
M
Moat
Routing telemetry that continuously improves per-week winner picks.
Risk
Vendor terms and API churn; provider bundles ship the routing.
OPPORTUNITY 03

Agent reliability runbook workshop

What
Training for Thai teams on eval gates, checkpoints and agent QA rituals; includes installing the monitor.
Who pays
Companies pay per seat for risk-managed agent adoption.
Price
฿2,000–6,000/seat — hypothesis
Why now
Teams adopting agents hit the reliability wall the post describes.
Effort
S
Moat
Case library from real Thai deployments.
Risk
Niche willingness to pay for process training.