The news
Stack Overflow's latest pulse survey reports that AI agent usage in professional settings has nearly doubled year over year — jumping from 31% to 59%. Despite the surge, most deployments remain single-agent and human-monitored, not the autonomous multi-agent orchestration that dominates the vendor pitch circuit.
Our take
This data matches what's actually happening in practice. Teams are experimenting with agents. Very few are trusting them to run unsupervised.
That's not a failure state. That's the correct answer for where most organizations actually are.
Here's the pattern that repeats: a team gets excited about agents, spins one up to handle something like lead routing or follow-up sequencing, and then discovers the agent makes confident mistakes — wrong contact, wrong stage, wrong message — and nobody catches it until the damage is done. The fix isn't better AI. The fix is a human checkpoint in the workflow. Which is exactly what the majority of respondents said they have.
The "agents on a leash" framing undersells what's actually happening. Human-in-the-loop isn't a limitation — it's how you build the trust that eventually earns the agent more autonomy. The teams that will successfully run multi-agent systems in 18 months are the ones using monitored single agents today. They're building the observability habits, the failure taxonomies, the process documentation that makes expanded autonomy safe.
The GTM-specific risk here is real. Agents operating on your CRM, your outbound sequences, or your campaign logic have direct revenue impact. A hallucinating agent that updates Salesforce fields, fires a nurture sequence to the wrong segment, or scores a lead incorrectly doesn't just waste time — it corrupts the data layer that every downstream decision depends on.
Single agent, human monitored, one clear job: that's not a consolation prize. That's the blueprint.
So now what?
- Audit what your agent is actually touching. If it has write access to your CRM or marketing automation platform, you need a review step before it commits anything — full stop.
- Define the failure mode before you deploy. What does a wrong answer look like? What's the blast radius? Document it. Your future self will thank you.
- Log everything. Even if you're not reviewing every output now, you want the history when something goes sideways.
The teams getting the most from agents aren't the ones moving fastest — they're the ones who know exactly where their agents are allowed to go.
---