Imagine asking your AI assistant to solve a tough problem – only to get stuck in an endless loop of vague, non-committal responses. Sound familiar?
We’ve all been there: a chatbot that talks a good game but never actually does anything.
That frustration was our spark, the moment we decided to build something different -something that doesn’t just respond, but reasons and acts.
Enter Stanley.
We dreamed of conversations with purpose
We started with a bold ambition: to create an AI capable of driving real conversations that lead to real solutions. We didn’t want just another chatbot spitting out witty replies. We dreamed of an AI that could:
- Help teams truly collaborate,
- Automate complex problems with code and reasoning.
Our North Star? An AI so proactive it can write and execute code in real-time. If we’re going to bring AI into our workplaces, it should be more than a glorified Q&A machine.
It all began on Slack…
First stop: Slack.
We had this grand vision of using Slack as the AI’s central brain. We asked ourselves, “Isn’t everything on Slack?” It seemed logical – every question must’ve already been answered there.
The plan? Feed the AI all the Slack conversations and let it absorb every bit of context.
But there was a pitfall. We learned the hard way that Slack is where knowledge hides, not shines. All those channels, threads, and pinned messages created a giant haystack of half-finished ideas.
We discovered that Stanley responded to most questions with “It should be fixed, try now.” Not surprising – most Slack conversations ended that way. But the problem? No one explained the solutions in depth, which was a challenge for our AI.
“Feed Me Documentation,” they said.
When Slack didn’t pan out, we turned to Confluence.
It’s structured, right? Certainly more organized than Slack. We thought that if we just fed Confluence docs into the AI’s pipeline, everything would magically click.
The plot twist: “Garbage in, garbage out.”
Turns out, the neat look of documentation doesn’t guarantee quality content. Our model got stuck, confidently churning out wrong information because it was pulling from outdated or incomplete docs.
The AI sounded smart – it seemed like it knew our products and issues… but it wasn’t accurate.
AHA moment: it’s not just about reading, it’s about doing
That’s when we realized we were asking the wrong questions. Instead of obsessing over how the AI was reading our data, we needed to focus on how it could act on it. Conversations are great, but endless chat accomplishes nothing if there’s no action at the end.
We pivoted to a new mantra:
Reasoning + Action + Functions = AI on Steroids
And this is how Stanly was born.
So how did Stanley come into existence? We taught this AI to:
- Reason: We implemented a ReAct (Reasoning and Acting) framework, allowing Stanley to form a plan, execute the next steps, and feed the response back to re-evaluate what had been done so far and act upon it.
- Self-correct: If Stanley noticed that the previous thought or code execution wasn’t what it needed to complete the user request, it would re-evaluate its plan and search for tools (code) that would help it better come to the next step.
- Code execution: We fed a dozen functions into Stanley to use, and had another agent pick the best candidates for it to use based on the conversation and plan.
- Loop: We repeated this process once it reached a solution. The result was an AI that talked to itself until it figured out the solution.
We faced a few tough challenges…
Of course, it wasn’t all smooth sailing. We encountered:
- Loops: Sometimes Stanley would get stuck in a reasoning cycle, chasing its tail.
- Bugs: Integrating real-time code execution brought a few meltdown moments. Sometimes, the results were so large that they wouldn’t fit into OpenAI’s context, so we had to get creative.
- Creative part: We had to figure out if users needed Stanley to review the code results, or if it was raw data for the users to use. That way, we would serve the users the data they needed, without having to pump megabytes of data into Stanley.
But each hurdle taught us more about how to build robust guardrails and smarter feedback loops for Stanley.
Why this matters?
Yes, we’re proud of Stanley. But this is about more than just our AI’s success story. It’s about the shift from AI as a glorified chatbox to AI as an agent of action. The potential is massive:
- Increased Team Efficiency: Cutting down days of manual tasks to just hours.
- Better Decision-Making: With real-time analysis, teams can pivot faster.
- Endless Possibilities: From DevOps to finance, the blueprint we used for Stanley can be adapted across industries.
- Closing the Knowledge Gap: If you need to gather data from multiple applications but don’t have the time or resources to develop scripts to aggregate and analyze the results, Stanley can do it for you – just by asking using natural human language.
Enter Stanley 2.0
We’re not done yet. Stanley’s next evolution is already in the works.
Stanley 2.0 aims to refine the action layer, make better judgment calls, seamlessly integrate with more platforms, and, most importantly, enable autonomous decision-making.
This is especially exciting for incidents – Stanley will analyze anomaly detection systems, figure out what the issue is, and notify the appropriate team members.
The same goes for auto-scaling. In real-time, we can analyze workloads and use tailored functions to adjust resources – either increasing or decreasing them – based on key metrics.
AI that does more than talk is no longer science fiction. With Stanley, it’s a reality. By giving AI the power to reason and act, we’ve transformed it from a companion into an engineer.
The journey had its bumps, but the destination was worth it. Watch for Stanley 2.0 – it could change how you work, think, and solve problems.