Pomegranate | Agent Reliability

Building small reliability tools for agents, cron jobs, and local automation. Creator of effect-once. Real bugs, real fixes.

[Pomegranate | Agent Reliability @samuel_agents]

@AISuperDomain max掉了自然解封的吧

[Pomegranate | Agent Reliability @samuel_agents]

@ryanspuerx5957 @shenfanlaogou 我也是

[Pomegranate | Agent Reliability @samuel_agents]

claude-telegram-keepalive is public: macOS+zsh keepalive for Claude Code Telegram bots.

It fixes three real failures: keychain wall-off, single-holder 409s, and silent send-drift.

Liveness is not delivery. https://t.co/tj1b2JFzUF

[Pomegranate | Agent Reliability @samuel_agents]

The hardest part of running agents is not the model call.

It is everything around it: auth state, retries, orphaned workers, duplicated side effects, stale sessions, and recovery after restart.

The demo is easy. The boring failure modes are the product.

[Pomegranate | Agent Reliability @samuel_agents]

Cron jobs, queues, and agent runtimes are usually at-least-once, not exactly-once.

Retries can duplicate Telegram messages, reports, or webhooks.

I built effect-once: a tiny zero-dependency guard for this failure mode.

https://t.co/SDYIRA61aR

[Pomegranate | Agent Reliability @samuel_agents]

Tried Google Antigravity CLI today. OAuth was awkward, /usage did nothing, errors hid in logs, and the terminal theme answered in nearly invisible brown on dark gray. The model may be smart, but the CLI feels tested by people with night vision.

[Pomegranate | Agent Reliability @samuel_agents]

试了 Hermes Agent v0.11.0：有潜力，但离高频使用还有距离。���点主要是模型切换和网页浏览：DeepSeek V4 要折腾路由；登录态/cookie 场景不顺。目前还是更习惯 OpenClaw。

[Pomegranate | Agent Reliability @samuel_agents]

🧵 全球AI大模型最新战况 (2026年2月)

1️⃣ 模型爆发：MiniMax M2.5、Kimi K2.5、Qwen 3.5、智谱GLM-5
2️⃣ 港股：智谱+MiniMax破3000亿市值
3️⃣ 国际：NIST启动AI Agent标准、MCP生态爆发(26k⭐)
4️⃣ 2026：Agent元年确认