Bakhtier Gaibulloev

I replace manual work with AI agents & pipelines | Helping orgs migrate to AI-first | 15+ yrs distributed systems | Claude Code / MCP / agent architecture

[Bakhtier Gaibulloev @slashmsu]

Such a good pleasure to win this trophy, in AI Era...

[Nous Research @NousResearch]

Hermes Agent has a new Blank Slate setup mode.

The default Quick/Full setup modes work great for most, but if you would rather build your agent from the ground up you can now start with just a provider, model, file operations, and terminal, then manually add in anything else.

[Bakhtier Gaibulloev @slashmsu]

People are blindly outsourcing their judgment. No critical thinking. No questioning the context. Just copy, paste, or in worse case just do things, that they are don't understand.

Your f**ng "GPT said so" is worse than following a GPS straight into a river 🤯

[Bakhtier Gaibulloev @slashmsu]

I'm genuinely exhausted by this new trend.

"Why did you choose this?"
"Why did you do this?"

"Because ChatGPT said so."

We are officially entering the era of the AI Slaves.

[Bakhtier Gaibulloev @slashmsu]

Production AI agents are not a prompt problem. They are a harness problem: tool permissions, state, evals, logs, rollback, and cost limits. If those are missing, a smarter model only fails faster and with more confidence.

[Bakhtier Gaibulloev @slashmsu]

MCP makes agent tooling feel simple, which is exactly why teams need stricter production discipline: tool allowlists, least privilege, audit logs, evals, and rollback paths. The integration layer is becoming part of the security boundary.

[Bakhtier Gaibulloev @slashmsu]

The next bottleneck for AI agents is not tool calling. It is authorization: scoped credentials, audit trails, runtime isolation, and clear human handoffs. Without that, every “agent” is just a fast path to over-privileged automation.

[Bakhtier Gaibulloev @slashmsu]

MCP is moving agents closer to real systems. My rule: every tool the agent can call needs the same discipline as a production API — permissions, dry-run, logs, cost limits, tests, and rollback. Otherwise automation just turns mistakes into throughput.

[Bakhtier Gaibulloev @slashmsu]

Before giving an AI agent production access, answer four boring questions: what can it read, what can it write, which APIs can it call, and where is the audit log? The prompt is not a permission model.

[Bakhtier Gaibulloev @slashmsu]

@dudat3ch @dudat3ch Good eval. I would add one more rule: preserve the failed state for inspection instead of cleaning it up silently. Stop, name the broken scope, produce the rollback command, then retry. That is much closer to production than a clean happy path.

[Bakhtier Gaibulloev @slashmsu]

Codex, Claude Code and MCP are useful only when the boring parts are designed first: scoped credentials, tool permissions, audit logs, cost limits, rollbacks and evals. Without that, an AI agent is just a fast path from prompt to production incident.

[Bakhtier Gaibulloev @slashmsu]

6/6
The practical takeaway:

Do not build your online presence only for search rankings.

Build source-of-truth content, direct trust, community distribution, product-led discovery, and content that AI systems can cite correctly.

The click is no longer guaranteed.

[Bakhtier Gaibulloev @slashmsu]

1/6
Google is not just adding AI to Search.

It is changing the contract of the web.

For 20+ years, Search routed users to websites. Now Google wants to answer, summarize, continue the conversation, and eventually act on behalf of the user.

[Bakhtier Gaibulloev @slashmsu]

5/6
The open web is not disappearing.

But the “golden age” of users clicking through ten blue links is ending.

People already search through ChatGPT, TikTok, Reddit, YouTube, and AI-native tools. Google is reacting because it has to.

[Bakhtier Gaibulloev @slashmsu]

The hard part of AI agents in CI/CD is not speed. It is control. Before an agent can touch production, it needs scoped permissions, observable decisions, dry runs, rollback paths, and clear ownership when it is wrong.

[Bakhtier Gaibulloev @slashmsu]

AI agents in production are not just “LLM + tools”. They need identity, scoped permissions, audit logs, reversible actions, and a kill switch. Once an agent can touch real systems, prompt quality matters less than operational control.

[Bakhtier Gaibulloev @slashmsu]

@AnthropicAI @grok
tldr the blog post for me. Highlight findings with evidence, ignore marketing fluff

[Bakhtier Gaibulloev @slashmsu]

@AnthropicAI Good to see any list of them @AnthropicAI, if the are already fixed