The agent observability gap: why most teams are flying blind
By The Lucidrail Team · 2026-07-09
Most stories about AI agents going wrong open with a dramatic incident — a runaway loop, a deleted database, a surprise bill. This one starts with something quieter and far more common: most teams running agents in production cannot see what those agents are doing at all.
According to zlatkov, who leads agent developer tools at Progress, only about one in five teams instrument observability for their agents from day one. Treat that as a directional signal rather than a hard statistic — it comes from a practitioner’s own read of the field, not a peer-reviewed study. But the shape of it rings true.
If roughly one in five teams can see, then four in five are flying blind. And the ones who can see mostly built the tooling themselves, because until recently nothing sat on the shelf ready to buy.
Four in five are flying blind
“Flying blind” is not an exaggeration here. It means there is no live view of what each agent is doing, how much it is spending, or whether it has quietly gone off the rails. The agent runs. The work happens, or it doesn’t. And the operator finds out which one it was after the fact — usually when something outside forces the issue.
That is a strange way to run software that can spend money, send messages, and change data on its own. Nobody would ship a payments service with no dashboard and no alerts. Yet that is roughly the default posture for agents today.
What blind actually looks like
Being blind has three usual tells. You learn what went wrong when the bill arrives, when the wrong email has already sent, or when a row that should be there is gone. The signal is the damage.
A team at Artificial Lab put the cost version of this plainly. They ran seventeen agents and had no running view of the spend at all:
One day we realized we had no idea what they were spending… The UI for all of that spending was our credit card statement. Once a month. After the fact. — peiyaooo, on running seventeen agents at Artificial Lab (Show HN)
Once a month, after the fact, is not visibility. It is a receipt. By the time you read it, whatever happened has already happened, and the money is already gone.
The tell: the people who can see built it themselves
Here is the strongest evidence that this is a real, structural gap and not just a few unlucky teams. The operators who do have visibility mostly did not buy it. They built it — one weekend at a time — because nothing off the shelf did the job.
One developer running several coding agents at once described the everyday version of the problem, then went and built a tool to fix it:
There’s no good way to see which of your 8 running sessions actually needs attention. I was losing work and wasting money. — Beefin, after running several coding agents at once (Show HN)
Another, running agents for an e-commerce store, could not even reconstruct what his agents had said to customers after they started going wrong:
When I tried to figure out what happened, I had scattered logs across services with no way to prove what the agents actually said. — shotwellj, on building a tamper-evident audit trail for agents (Show HN)
A home-made spend dashboard here, an audit “black box” there, a session multiplexer somewhere else. When many capable people independently rebuild the same missing layer, that layer is not a nice-to-have. It is infrastructure that should already exist.
Blind is the default, not safe
The quiet lesson in all of this is about starting state. When you stand up a new agent, the default is not a safe, watched system that occasionally needs attention. The default is no visibility at all. Safety is the thing you have to add — and most teams have not added it yet.
So the first question for anyone running agents is not “are my agents well-behaved?” It is simpler: can I see them at all? Everything else — caps, approvals, an audit trail — depends on that one answer being yes.
This is the gap we started Lucidrail to close. One console shows every agent in your fleet, what each is doing and spending, live — not next month on a statement. Under it sit per-company budgets with caps that stop a runaway before it bills, and an audit trail where every action traces back to the goal that prompted it. Visibility you switch on, instead of a dashboard you build.