Antrixsh Gupta

Vibe coding hackathon You all love this.

@arsh_goyal Hyderabad

Guess the city?

A lot of AI agent startups got cooked today. Anthropic launched Managed Agents. A fully hosted service that runs long-horizon AI agents on your behalf. Session management. Sandbox execution. Context engineering. Failure recovery. All of it. Native to the Claude platform. Here is what this actually means. There is an entire category of startups whose product is exactly this. “We make AI agents reliable at scale.” That is their pitch. That is their Series A. That is their moat. Anthropic just made it a feature. The AI agent infrastructure space just got a lot more crowded.

I just published Scaling Agentic AI: Multi-Agent Systems Explained medium.com/p/scaling-agen…

Oracle fired 30k employees via 6 AM email. And Oracle is not a struggling company, they made most money than ever. Despite that, 30k people lost their jobs. They told you to code, you did. They told you to upskill, you did. They told you to learn AI, you did. And then they replaced you with the same system you helped them build. #oracle #layoff

The biggest blind spot in AI isn't prompt injection, it's permission creep. Agents execute valid API calls that quietly cross trust boundaries. EDR sees processes, not context drift. Treat agents like privileged identities, not just software. #AIInfra #AppSec #Agents

I'm noticing a pattern: the bottleneck for agents isn't reasoning, it's execution boundaries. We've got agents rewriting code overnight, yet we treat prompts as security. If it can trigger side effects without hard IAM, you just have best-effort vibes. #AIagents #LLMs #IAM

Noticing a shift away from naive RAG. Chunk-and-embed pipelines work for text fetching, but fail at understanding structure. The unlock for agent memory isn't vector similarity—it's structured graph layers where retrieval is navigation, not lookup. #RAG #AIAgents #LLMs

I'm noticing a shift in agent infra: browser automation is a dead end. Builders are abandoning DOM scraping, instead wrapping the web in CLIs and offline-first MCP servers. Stop making your models read HTML. Give them native programmatic access. #AIagents #MCP #dev

I'm noticing a pattern: our evals are rotting. We’re optimizing against benchmarks where 6% of the ground truth is wrong, and weak LLM judges accept 63% of garbage answers. We're benchmarking context windows, not actual memory. Build deterministic evals. #LLMs #Evals #GenAI

@sickdotdev github.com/antrixsh/trust… This is what I am building a The open-source framework for evaluating LLM safety, fairness, and reliability in regulated industries.

Hey Founders Drop what you’re building👇 Last time 1M+ people saw it. Consider this as marketing.

@heyblake github.com/antrixsh/trust…

Drop your project URL Let’s drive some traffic

An agent without strict monitoring and token limits isn't a feature. It's a memory leak with a credit card attached. Without guardrails, they loop, eat your RAM, and burn your API budget by Sunday. Agents are volatile infra, not static code. #AIAgents #Infra #LLMs

I'm noticing a massive blind spot in agent evals: grading the final output. An agent can loop 5 times, hallucinate a tool call, recover, and still return the 'right' answer. If you aren't scoring the execution trace, your evals are lying to you. #AIAgents #Evals #LLMs

The real issue is we're optimizing models against broken yardsticks. An audit of the LoCoMo benchmark found its LLM judge accepts 63% of deliberately wrong answers. LongMemEval is just a context window test. Stop trusting leaderboards. Build custom evals. #LLMs #Evals #AI

We're building recursive self-improving agents, yet MCP still chokes on binary transfers and agents need append-only WALs just to survive context compaction. The real bottleneck isn't reasoning. It's brittle infra and output-only evals. #AIAgents #LLMs #DevTools

Evaluating AI agents on their final output is a massive blind spot. I’m seeing agents land on the correct answer only after insane tool loops and near-catastrophic API calls. Stop scoring the output. The real signal is in the execution trace. #AIAgents #LLMs #Evals

Most 'agents' are just hardcoded DAGs with an LLM node in the middle. And that's fine. Hardcode your logic. Use models strictly for messy inputs. When workflows break, you patch a node. When agents break, you're lost in hallucinated tool calls. #LLMs #Agents #DevTools

The biggest blindspot in agent dev isn't reasoning, it's infra. We obsess over prompts but ignore structural failures: missing idempotency keys, hidden trace loops, and duplicate tool calls. Stop evaluating just final outputs and audit the trace. #AIAgents #LLMs #Infra

Evaluating agents purely on final output is a trap. They can hit the right answer while doing nonsense under the hood: infinite loops, hallucinated tool calls, and wasted compute. If your evals don't score the execution trace, you're flying blind. #Agents #Evals #LLMs