1/6
SRE did not make servers smarter. It made them trustworthy enough to run the internet.
We need the same discipline for agents.
Smarter models will not fix production failures. In critical ways they make agents less reliable: guardrails become obstacles, and the goal is what the model actually optimizes for.
trustabl.ai/blog/smarter-mod…
1
93
Satya seems very aligned with the discipline of Agent Reliability Engineering spelled out at agent reliability engineering.org
22
1/3 As models get smarter, they break more guardrails and other constraints. The future is sandboxes (AKA jails) like OpenShell. bbc.com/news/articles/c6vgy0…
1
1
33
2/3 Just a reminder that this and the Hugging Face hack are not anomalies, they are a trend... trustabl.ai/blog/smarter-mod…
1
16
3/3 This is why we desperately need an Agent Reliability Engineering (ARE) discipline, just like SRE. AgentReliabilityEngineering.…
10
5/6
If you are shipping agents that take actions — tools, money, tickets, infra, customer data — come help write the discipline.
Site + working group: agentreliabilityengineering.…
Commentable draft: docs.google.com/document/d/1…
1
10
6/6
People who should stress-test this, not applaud it:
@HamelHusain @sh_reya @eugeneyan @aparnadhinak @marcklingen @simonw @mipsytipsy @copyconstruct @hwchase17 @chipro @swyx @jxnlco @steverab @natebjones
If ARE is missing a tenet you already learned the hard way, say so on the draft.
1
24
Smarter models reduce agent reliability. trustabl.ai/blog/smarter-mod…
Does Agent Reliability Engineering grow into a discipline like Site Reliability Engineering I'm betting it does. @karpathy
github.com/mikehogan-thought…
1
35
Even Kimi K’s founder admits it:
Powerful models aren’t enough. Most agents are configured to fail.
Input validation, retry logic, observability, guardrails, tool reachability… skip these and you’re doomed no matter how smart the model is.
That’s exactly why we built Trustabl — open-source, local-first, deterministic hardening for your agents. Finds the breaks. Fixes them. Cuts token spend.
Watch Moonshot CEO explain 👇
Try it free → trustabl.ai + github.com/trustabl
Are your agents ready for production yet? 🔥
#AI #AIagents #KimiK #AgenticAI #OpenSource
1
2
106
I'll believe @realDonaldTrump election claims when the guilty individuals are convicted in open court. #ProveIt
1
36
My orphan granddaughter has a 24-hour stomach flu, throwing up. @SouthwestAir won't switch their flight to tomorrow, would rather get the whole plane sick... Not a fan of the new ownership. #Fail
No supervisor to talk to, so much for status and loyalty #SWAFail
Some argue the NYC mayor Mamdani is freezing rent because he doesn't understand economics. They claim he doesn't understand that maintenance, mobility, and availability drop like a stone and then landlords abandon money losing properties by the thousands, like they did in NYC during the 70s and 80s. What if he understands exactly what he's doing and plans to demonize landlords for poor maintenance the whole time. Then the government takes over the abandoned buildings "for the people". It's the perfect communist playbook.
43
Mike Hogan: Software Entrepreneur retweeted
As more agents go into production, reliability has become a real problem.
Missing timeouts, unguarded shell tools, SSRF exposure, missing observability, and agents with zero guardrails are now shipping to production.
Here’s the easy button solution.
Agents require context, just like AI models github.com/trustabl
@natebjones @fireship_dev @karpathy @mckaywrigley @mattshumer_ @AndrewYNg
1
4
187
A Simple Automated Way to Make Your Agents More Reliable
As more agents go into production, agent reliability has surfaced as a serious issue that requires immediate attention. Here’s your “easy button” solution. Trustabl is a static analyzer that scans agent repos built on Claude, OpenAI, Google ADK, and MCP and finds reliability and safety issues before they reach production. Missing timeouts. Unguarded shell tools. SSRF exposure. Missing observability hooks. Agents with no guardrails wired. And more…
The part that surprised people when we shipped it: fixing those issues is now almost entirely automatic.
Here's how it works end-to-end:
→ In your editor (Cursor / Claude Code): Connect the @Trustablai MCP server. Your AI coding assistant can query findings directly and implement fixes inline, no context switching, no separate tool.
→ In CI: Add two lines of YAML. The GitHub Action scans every PR, posts a readiness score as a sticky comment, and surfaces findings as inline annotations on the changed lines. Add your Anthropic API key and set auto-enrich: true. High-confidence fixes are applied and a fix PR is opened for your review. You accept what looks right.
→ For DevSecOps: Findings upload to GitHub Code Scanning via SARIF. They appear in the Security tab alongside your other supply chain signals. Same dashboard, same workflows, same alert routing. No new tooling to learn.
One scan. Three surfaces. The issues get found, explained, and fixed, mostly without you having to touch them.
We're open source (Apache-2.0) and free to start. If you're shipping agents, we'd love to know what you find.
github.com/trustabl
1
2
202






