How We've Been Building Agents That Actually Work

Hiya 👋

Everyone's got access to AI tools now but the companies seeing real results are doing a things differently than most.

This week I pulled together what's working: simple agents that actually help, why some AI projects take off while others stall, and what 1,200 workers told us they want.

Let’s jump in👇

How to Think About Agents (minus the hype)

Most companies are building AI agents wrong. They're creating overcomplicated Swiss army knives that break, or they don't understand what agents actually do for them.

After building agents that actually work, here's what we've learned: the best ones are simple, and most people completely misunderstand what they are.

What AI Agents Actually Are (And Why Most People Get It Wrong)

Think of AI agents like having a really capable assistant who knows exactly which tools to use for each part of a job.

Say you need to prepare for a client meeting tomorrow.

Without an agent, you'd manually search their website, check your CRM for past interactions, pull their latest financial reports, update your proposal deck, and schedule follow-up.

Five different tools, five different logins, lots of copy-pasting between them.

With an agent that has tools, you just say "prepare me for the Johnson Industries meeting tomorrow."

The agent searches the web for their recent news, queries your CRM database, fetches their financial data, updates your presentation template, and blocks time on your calendar.

Same tools, same results, but the agent knows which tool to grab for each step and connects the outputs together.

Good agents aren’t a Rube Goldberg machine. 

The reliable pattern most teams land on is simple: a loop that thinks, uses a tool, checks results, and repeats.

Picture: Input → LLM → (call a tool if needed) → Output → loop. That’s it, the diagram bellow literally shows this “while loop with tools,” and it’s the backbone used by popular systems like Claude Code and OpenAI Agents.

Why it matters: Simple beats clever. The more layers you add (multi-agent planners, complex graphs), the more things break. Teams that ship focus on three levers: good tools, good context, and basic evaluation, not fancy orchestration.

🧪 How we build them at Switchboard (in plain English)

We often use N8N to build agents and each agent has a purpose. We try to think of them in simple terms and name them.

For example, on a recent architecture-industry project, we needed early signals of upcoming capital projects before they get to RFP. The details are buried in city council and university board meeting minutes & transcripts.

The data was messy, unstructured, and scattered across dozens of sites.

We solved it with 4 agents, each with their own purpose and tool:

• The Watchman → a crawler checks source sites and uses one tool to grab and log new entries (PDFs, HTML, etc.)

• The Librarian → tracks sources, documents, what’s new, and what’s been scanned.

• The Brain → LLM reads documents, extracts opportunities, summarizes, and routes to the right people.

• The Town Crier: LLM distills and sends a digest to the business-development team once opportunities cross a threshold.

  
  Cost to run? About $1.65 in AI API model calls per weekly run across all the sources. This previously would have taken multiple people, multiple hours a week to do.

What mattered most wasn’t a fancy graph of agents, it was clear prompts, tight tools, guardrails, and iteration. We started around 60–70% accuracy and tuned to 90%+ by testing, adjusting prompts, and cleaning outputs.

👣 Start simple (a non-technical recipe)

1. Pick one job. “Scan these 10 sources weekly and flag BD opportunities.”

2. Map the steps like a person would. Where do you look? What matters? What do you ignore? Have a subject-matter expert record a 10–15 minute Loom talking about their thinking process out loud; turn that into plain-language rules and prompts for the AI agent.

3. Give it 2–4 tools max. Example: fetch_page, extract_text, classify_opportunity, notify_owner. Avoid Swiss-army-knife tools.

4. Evaluate + iterate. Run 10 docs at a time, compare to an expert doing it, nudge prompts, repeat until it’s boringly consistent.

Braintrust wrote a great article on how to think about building Agents →

Mark Cuban recently told Fortune that most companies “don’t understand how to implement” AI, and that Gen Z’s biggest opportunity is helping businesses figure it out. He’s right.

Cuban’s angle: it’s not about building new AI tools. It’s about walking into a business, spotting where AI can help, and customizing it to real workflows. For small and mid-sized companies without AI budgets or expertise, that skillset will soon be as essential as knowing email or Excel.

Here’s where I agree, and where I think the data backs him up. MIT just reported that 95% of AI projects don’t deliver ROI. And it’s not because the tech doesn’t work. It’s because too many leaders still treat AI like plug-and-play software instead of a process partner.

What we’re seeing at Switchboard:

• Fix the back office first → AI shines when it replaces bottlenecks and BPO contracts, not when it’s hyped in front-office demos.

• Start small, scale later → Success comes from launching one meaningful workflow, not boiling the ocean with endless roadmaps.

• Partner with specialists → External projects succeed 67% of the time vs 30% in-house.

The real challenge isn’t the tech, it’s the people, the workflows, and the willingness to change how work gets done. That’s why 95% fail. And it’s also why the 5% who get it right are unlocking serious ROI.

🎧 If you want the deeper dive, Tom and I unpacked this in the latest Work Less Podcast episode.

Kyla Scanlon, bestselling author of In This Economy? and a recent guest on The Daily Show, has a rare gift for making modern work and economics accessible. Her latest project: a survey of 1,200 workers across industries, asking how they feel about AI.

Why it matters: Most AI surveys focus on executives. Kyla flipped it: she asked the people on the frontlines of implementation. The result is a grounded snapshot of how workers are navigating AI’s messy middle.

👉 Key takeaways:

• AI hopes: Workers want it to handle the boring parts, such as paperwork and repetitive admin, so they can focus on meaningful work.

• Concerns: Not mass layoffs, but erosion of career opportunities, creativity, and human connection (especially in healthcare, education, and creative fields).

• Trust gap: Most workers “somewhat trust” employers on AI. No industry reported majority “complete trust.” 62% want shared decision-making in how AI gets deployed.

• Training gap: Only ~60% have received AI training; rates are much lower in creative fields (20%) and entertainment (5%).

• Policy asks: Training & upskilling funds, algorithmic transparency, and safety nets for displaced workers.

The big picture: As Kyla put it on X, this isn’t just about efficiency or job loss. Workers are asking: What makes work meaningful? What parts of being human do we want to preserve?

Know someone who’d enjoy this newsletter?
Forward it along, or send us a note if there’s a topic you’d like us to unpack next.