StackAI retweeted
StackAI launched Multi-Agent Teams! ๐ค
Here's the problem we built it to solve ๐
A customer's agent works well when asked to do five tasks. At fifteen it starts getting unreliable. At thirty it's guessing.
Every time you widen the scope, you expand the decision space the model searches on every single call, and every new instruction competes with the ones already there. You're not making the agent smarterโฆ you're making the problem harder.
Now, instead of one model doing everything, an Orchestration system routes each job to a Sub-agent that only knows what it needs to know: its own instructions, its own tools, its own knowledge bases. Narrow scope, fewer wrong turns. They run in parallel, and the Orchestrator brings the results back as one accurate output.
Same models. Same tools. Better architecture.
StackAI retweeted
One of my favorite agentic workflows... Please never organize documents by yourself again lol
We now use @stackai instead:ย
๐งEvery time a new document arrives via an email
๐ค An AI Agent automatically uploads it and renames it in SharePoint/Gdrive
๐ and follows the correct folder path, metadata, and naming conventions!ย
Thatโs how agentic AI should feel. Running seamlessly in the background & proactively taking work off of your plate.
StackAI retweeted
Iโm thrilled to share that StackAI is joining forces with Asana.
From the moment we met Dan Rogers, Asanaโs CEO, Arnab Bose, Asanaโs CPO, and the rest of the team, it was clear they deeply understood what makes StackAI different: how we build, who we serve, and why it matters.
When we started StackAI in early 2023, we believed LLMs would fundamentally redefine how work gets done. Today, with the rise of AI agents, that future is arriving faster than anyone expected; and we believe weโre uniquely positioned to help shape it.
Over the past few years, companies across a wide range of industries have adopted StackAI to build reliable AI workflows for critical business operations. That momentum exists because of an extraordinary team that has obsessed over every detail of the product experience: from workflow reliability and orchestration to scalability, governance, and usability.
At the same time, Asana has spent nearly two decades building one of the most trusted platforms for work management, used by 85% of the Fortune 100 and loved by millions of users worldwide.
What makes this partnership especially exciting is our shared belief that the future of work will be built around seamless collaboration between humans and AI agents. Together, we have an opportunity to build something entirely new: a true operating system for human-agent execution across teams, tools, and data.
Iโll continue leading the StackAI team alongside my co-founder Bernard and our incredible team, reporting directly to Arnab Bose. Arnab brings a rare combination of ambition, humility, and product leadership, and I couldnโt imagine a better partner for this next chapter.
Importantly, StackAI remains StackAI: same team, same product, same brand, same commitment to customers. What changes is our ability to move faster, scale further, and bring more powerful AI capabilities to organizations around the world as part of Asana.
This is not just about making individuals more productive. Itโs about enabling entire organizations to operate differently: with humans and AI agents working side by side through governed, reliable workflows.
To our customers: thank you for trusting us early. We plan to keep shipping faster than ever.
To our team: this moment belongs to you. Your craftsmanship, intensity, resilience, and ambition made all of this possible. Now we get to build at an entirely different scale.
Excited for whatโs ahead, letโs keep cooking!
StackAI retweeted
One question I hear constantly while filming inside AI startups is: what happens when OpenAI just builds this?
While filming INSIDE @stackai, co-founder @BernAceituno gave a pretty compelling answer.
StackAI has raised $16.6M to build an enterprise AI agent platform that lets organisations automate complex workflows without writing code.
For hospitals, defense contractors, and financial institutions, AI infrastructure has to meet strict compliance requirements. Once those workflows are embedded into an organisation, the switching cost isnโt the model, itโs everything built on top of it.
The companies that survive probably wonโt just be the ones with the best models. Theyโll be the ones that enterprises cannot remove.
Their INSIDE Startups episode crossed 30,000 views within 48 hours and drove over 500 high-intent clicks through to their website from YouTube alone. It's well past that now. Full episode link in comments!
StackAI retweeted
For a company automating enterprise workflows, @StackAI couldnโt let their website become a bottleneck. After building their site in Framer in just two weeks, their lean team now ships site updates daily and has scaled to nearly 2,000 pages, all without relying on developers.
StackAI retweeted
Everyone go home, @stackai just achieved AGIโฆ
Just kidding ;)
But this is getting very close.
Agentic Workflows... 100x more powerful.
StackAI retweeted
We at @stackai got early access to GPT-5.4 โ and the results are remarkable.
One of our benchmarks (analyzing financial records across 50,000 pages) had never been passed by any LLM. GPT-5.4 is the first to clear it.
Impressive work, @OpenAI ๐ฎ
StackAI retweeted
95% of AI agent demos never make it to production.
Yet 79% of enterprises expect full-scale agentic AI adoption within three years.
So what's the disconnect?
Most companies jump into AI agents without understanding what makes them fail at scale. The gap between demo and production is massive.
Weโve created this free guide with @stackai and @weaviate_io that breaks down exactly what goes wrong:
๐ญ. ๐ฆ๐ฒ๐ฐ๐๐ฟ๐ถ๐๐ ๐ฎ๐ป๐ฑ ๐ด๐ผ๐๐ฒ๐ฟ๐ป๐ฎ๐ป๐ฐ๐ฒ: Why agents leak data without proper access controls
๐ฎ. ๐ฅ๐ฒ๐๐ฟ๐ถ๐ฒ๐๐ฎ๐น ๐พ๐๐ฎ๐น๐ถ๐๐: How poor RAG implementation causes hallucinations
๐ฏ. ๐๐๐ฎ๐ฟ๐ฑ๐ฟ๐ฎ๐ถ๐น๐ ๐ฎ๐ป๐ฑ ๐ฒ๐๐ฎ๐น๐: The protection mechanisms that keep agents reliable
๐ฐ. ๐ฆ๐ฐ๐ฎ๐น๐ถ๐ป๐ด ๐ฐ๐ต๐ฎ๐น๐น๐ฒ๐ป๐ด๐ฒ๐: Why complexity grows nonlinearly with multi-agent systems
Plus, real-world use cases showing how to build production-grade agentic RAG systems.
Get your free copy here ๐ stack-ai.com/whitepaper/weavโฆ
StackAI retweeted
Think RAG is just vector search and retrieval?
It's actually 7+ different architectures (you might be using the wrong one)
1๏ธโฃย ๐ก๐ฎ๐ถ๐๐ฒ ๐ฅ๐๐ - The Vanilla approach. Documents get chunked, embedded, and stored in a vector database. When a query comes in, you retrieve the most similar chunks and pass them to the LLM.
2๏ธโฃย ๐ฅ๐ฒ๐๐ฟ๐ถ๐ฒ๐๐ฒ-๐ฎ๐ป๐ฑ-๐ฅ๐ฒ๐ฟ๐ฎ๐ป๐ธ - Naive RAG + a crucial step: after initial retrieval, a reranker model re-scores and reorders the results for actual relevance. This catches cases where semantic similarity doesn't perfectly align with what the user actually needs.
3๏ธโฃย ๐ ๐๐น๐๐ถ๐บ๐ผ๐ฑ๐ฎ๐น ๐ฅ๐๐ - Handles more than just text. Images, videos, audio - this architecture uses multimodal embedding models to encode different data types into the same vector space, then retrieves and generates responses across modalities.
4๏ธโฃย ๐๐ฟ๐ฎ๐ฝ๐ต ๐ฅ๐๐ - Instead of treating documents as isolated chunks, this approach builds a knowledge graph that captures relationships between entities and concepts.
5๏ธโฃย ๐๐๐ฏ๐ฟ๐ถ๐ฑ ๐ฅ๐๐ - Combines Vector Search with Graph RAG. By combining semantic retrieval with structured relationship mapping, you get a system that understands both the "what" (intent) and the "how" (connectivity) of your data.
6๏ธโฃย ๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐ฅ๐๐ (๐ฅ๐ผ๐๐๐ฒ๐ฟ) - Instead of a single retrieval path, an AI agent decides which search engine or knowledge source to query based on the user's question. It might hit a vector database for one query, a web search for another, or multiple sources and combine them intelligently.
7๏ธโฃย ๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐ฅ๐๐ (๐ ๐๐น๐๐ถ-๐๐ด๐ฒ๐ป๐ ๐ฅ๐๐) - The most sophisticated. Multiple specialized agents work together, each with access to different tools and databases. One agent might search internal docs, another queries external APIs, a third handles web search - all coordinating to answer complex queries that require information from multiple domains.
The architectures get progressively more powerful but also more complex to implement and maintain. Start simple, then level up as your use case demands it.
This was just a peek into @stackai and @weaviate_io latest ebook about building production-grade agentic RAG systems, get your free copy here: stack-ai.com/whitepaper/weavโฆ
StackAI retweeted
The difference between a cool proof-of-concept and a system you'd actually trust with customer data? ๐๐๐ฎ๐ฟ๐ฑ๐ฟ๐ฎ๐ถ๐น๐.
If youโre still treating guardrails as an afterthought, listen to this:
When enterprises deploy AI agents to handle real workflows like compliance reviews, claims processing, and RFP drafting, there's zero margin for hallucination. A chatbot making up facts is annoying. An agent fabricating regulatory guidance or policy language? That's a compliance nightmare.
So how do you build agents that fail safely instead of confidently wrong? Here are 4 guardrail implementations:
1๏ธโฃย ๐๐ฑ๐ฎ๐ฝ๐๐ถ๐๐ฒ ๐๐ฒ๐ฒ๐ฑ๐ฏ๐ฎ๐ฐ๐ธ ๐๐ผ๐ผ๐ฝ๐ โ The agent's Thought-Action-Observation cycle. After each action, the agent observes the outcome and decides: is the task complete? Do I need another tool? Should I ask for clarification? This continuous evaluation prevents agents from barreling forward with bad assumptions.
2๏ธโฃย ๐๐ผ๐ฟ๐ฟ๐ฒ๐ฐ๐๐ถ๐๐ฒ ๐๐ฐ๐๐ถ๐ผ๐ป โ Multi-stage evaluation where outputs pass through draft โ critique โ verify stages before finalization. Think of it as automated peer review. The agent doesn't just generate a response and ship it, it evaluates its own work against retrieved sources and quality criteria.
3๏ธโฃย ๐๐๐บ๐ฎ๐ป ๐ถ๐ป ๐๐ต๐ฒ ๐๐ผ๐ผ๐ฝ โ For high-stakes decisions, the agent routes drafts to human reviewers (via Teams, email, etc.) and waits for explicit approval before proceeding. Once the human responds, the workflow resumes automatically. You get the efficiency of automation with the accountability of human oversight.
4๏ธโฃย ๐๐บ๐ฒ๐ฟ๐ด๐ฒ๐ป๐ฐ๐ ๐ฆ๐๐ผ๐ฝ โ When the agent lacks sufficient context from authorized sources, it returns a clear no result state instead of improvising. This is probably the most important guardrail: teaching agents to say โI don't knowโ rather than hallucinate.
When these four guardrails work together, you get agents that are genuinely trustworthy, grounded in real sources, traceable in their reasoning, and safe enough for regulated environments.
If you're building enterprise AI without this kind of multi-layered guardrail architecture, you're building prototypes. Which is fine! But don't confuse prototypes with production systems.
Check out the full ebook on building reliable enterprise AI with agentic RAG from @stackai and @weaviate_io: stack-ai.com/whitepaper/weavโฆ
StackAI retweeted
Most companies will laugh you out of the room if you try to replace people with an unsupervised AI process.
AI + Humans > AI alone.
Every day of the week and twice on weekends.
Go out and talk to these companies: There are so many use cases you can't solve with AI alone.
The trick is to create a process that gives people control.
Something as simple as an approval step before an agent takes action, and many use cases will become doable.
Take a research agent as an example.
Here is what the agent does:
โข Looks into AI topics
โข Sends a summary of ideas to Slack
โข Waits for the user to approve those ideas
โข Writes the approved notes directly into Google Docs
Without the approval gate, the agent will litter the document with garbage. It would go off in random directions, and people would spend more time fixing its output than it would've taken them to do the work themselves.
With the gate, the user approves a direction, and the agent will execute it cleanly.
This pattern works everywhere:
โข Sales reviews leads before they hit HubSpot
โข Finance signs off before invoices get processed
โข DevOps confirms before a deploy runs
โข Legal checks before a contract is filled
StackAI retweeted
2025 was unbelievable for @stackai:
- Grew our revenue drastically (>11x since Series A).
- Saw what PMF feels like -> Revenue not just retain, but expands (sometimes 2x or 3x's).
- Released features our customers loved (auto-agents, on-premise, 1-click rag, browser use, etc...).
- Launched our deployments team. Now making customers expand and improve our platform.
Thatโs just the beginning. Now we will 10x down into what works. We are expanding across the board:
โ Marketing
๐ Sales
๐ ๏ธ Engineering
๐ป Deployments
Watch from the launchpad or jump on the rocket.
See openings below!
StackAI retweeted
Most enterprises waste months building agentic RAG systems for production.
If you start today, this framework gets you ready FAST:
Many enterprises have experimented with RAG; They've built chatbots, run pilots, and automated a few document lookups.
But in 2026 just experimenting is not enough. Production-ready AI requires moving past simple chatbots to โจAgentic RAGโจย
๐ช๐ต๐ฎ๐ ๐บ๐ฎ๐ธ๐ฒ๐ ๐ฅ๐๐ ๐ฎ๐ด๐ฒ๐ป๐๐ถ๐ฐ?
Instead of following a fixed sequence, agentic RAG uses AI agents that:
โ Route queries to specialized knowledge sources (vector databases, web search, internal APIs)
โ Validate retrieved context before using it
โ Make autonomous decisions about what information is needed next
โ Execute real actions across enterprise systems
๐๐ฒ๐ฟ๐ฒ'๐ ๐๐ต๐ฒ ๐ฑ-๐๐๐ฒ๐ฝ ๐ณ๐ฟ๐ฎ๐บ๐ฒ๐๐ผ๐ฟ๐ธ ๐๐ผ ๐ด๐ฒ๐ ๐๐๐ฎ๐ฟ๐๐ฒ๐ฑ:
1๏ธโฃย ๐๐๐ถ๐น๐ฑ ๐๐ฒ๐ฐ๐๐ฟ๐ถ๐๐ ๐ถ๐ป๐๐ผ ๐๐ต๐ฒ ๐ฎ๐ฟ๐ฐ๐ต๐ถ๐๐ฒ๐ฐ๐๐๐ฟ๐ฒ, ๐ป๐ผ๐ ๐๐ต๐ฒ ๐ฝ๐ฟ๐ผ๐บ๐ฝ๐
Start with platforms that enforce user authentication, role-based access control, and knowledge base permissions at the infrastructure level. On [StackAI] agents can only retrieve documents users are already authorized to see (no prompt engineering required).
2๏ธโฃย ๐๐ผ๐ฐ๐๐ ๐ผ๐ป ๐ฟ๐ฒ๐๐ฟ๐ถ๐ฒ๐๐ฎ๐น ๐พ๐๐ฎ๐น๐ถ๐๐ ๐ณ๐ถ๐ฟ๐๐
Break documents into machine-readable sections, enrich with metadata, and keep them synced with source systems. Weaviate's hybrid search combines semantic understanding with exact keyword and metadata filtering, so agents can reason over meaning while respecting policy boundaries.
3๏ธโฃย ๐๐ฒ๐๐ถ๐ด๐ป ๐ฝ๐ฟ๐ผ๐บ๐ฝ๐๐ ๐น๐ถ๐ธ๐ฒ ๐ผ๐ฝ๐ฒ๐ฟ๐ฎ๐๐ถ๐ป๐ด ๐ฟ๐๐น๐ฒ๐
Treat agent instructions as strict boundaries. Define what the agent is allowed to do, require consistent output formats, and explicitly instruct agents not to guess when information is missing.
4๏ธโฃย ๐๐บ๐ฝ๐น๐ฒ๐บ๐ฒ๐ป๐ ๐ด๐๐ฎ๐ฟ๐ฑ๐ฟ๐ฎ๐ถ๐น๐ ๐๐ต๐ฎ๐ ๐ฒ๐ป๐ณ๐ผ๐ฟ๐ฐ๐ฒ ๐ด๐ฟ๐ผ๐๐ป๐ฑ๐ถ๐ป๐ด
Every response should cite retrieved sources. High-trust systems often use multi-stage evaluation: draft, critique, verify. This mirrors human review while preserving automation.
5๏ธโฃย ๐ ๐ผ๐ป๐ถ๐๐ผ๐ฟ ๐ฐ๐ผ๐ป๐๐ถ๐ป๐๐ผ๐๐๐น๐
Track retrieval failures, hallucination risk, and output stability. Use LLM as a judge evaluation to measure which prompt patterns produce reliable results. Add human-in-the-loop for high-impact decisions.
The reality check is that Agentic RAG adds complexity and latency. But the alternative: AI that guesses instead of grounds, is a compliance nightmare in 2026.
Want to get started? Download @stackai and @weaviate_io's latest ebook on building with Agentic RAG here: stack-ai.com/whitepaper/weavโฆ
StackAI retweeted
Letโs reset the conversation on AI agents.
The term โAgentโ has become a catch-all for anything that calls a few APIs and runs an LLM loop (not very helpful). In reality, most so-called agents today are just chatbots with extra think steps. They talk extremely well, but few actually operate well.
So, what makes something an agent?
One-line answer: the ability to operate autonomously in complex environments to fulfill a task.
That work usually looks like this:
- Goal interpretation & planning โ turning vague requests into actionable plans, deciding if tasks need decomposition, which steps matter, and what to skip
- Information retrieval & tool execution โ knowing what it has, what it needs, picking the right tools, running them, collecting results
- Memory & context management โ tracking prior actions, intermediate results, failures, and constraints across steps
- Evaluation & adaptation โ checking progress, detecting failures, modifying approach when strategies aren't working
The distinction becomes obvious:
A chatbot tries to answer questions one-shot, while an agent runs a loop until it gets the job done by retrieving external info, tool execution, memory management, evaluation, etc.
Single-agent systems can handle such moderately complex work end-to-end, while multi-agent systems split responsibility across multiple specialist sub-agents.
The point here in all of this isnโt novelty or complexity. Itโs the shift from static workflows to systems that can reason over time.
And thatโs the definition worth anchoring on!
P.S. Want to learn more about building reliable, enterprise-ready agents? Download our free guide here: stack-ai.com/whitepaper/weavโฆ
StackAI retweeted
StackAI (@stackai) helps enterprises build custom AI Agents to automate dull, expensive work.
Companies like Nubank, BAE Systems, Morgan & Morgan, and Raiffeisen Bank use StackAI to automate compliance audits, legal research, HR, and customer support.
Congrats on the launch @BernAceituno & @tonirv!
ycombinator.com/launches/P7Tโฆ
StackAI retweeted
It only took 8 months, but here we have it...
RIP @StackAI_HQ, long live @stackai ๐
