There are two literatures about AI agents. The demo literature, where everything works. And the disillusioned literature, where someone let an agent loose on production and wrote a cautionary thread about it.
Both miss the point. Agents in production work extremely well — under specific conditions. We run them daily on our own operation and on client systems. These are the patterns that survive contact with reality.
What works
Agents inside guardrails, not agents with goals. The fantasy is “agent, grow my business.” The reality that works: “agent, every morning, read these inboxes, draft replies in this voice, flag anything about money or anger for a human.” Narrow scope, clear rules, defined output. The narrower the mandate, the higher the reliability — and you widen it gradually as trust accumulates.
Human approval gates at the blast radius. We sort every agent action by blast radius. Drafting is free — an agent can draft anything, because a bad draft costs nothing. Sending, publishing, deploying, spending: gated. The gate is cheap (one human glance) and it converts “terrifying automation” into “tireless assistant”. Most gates loosen over time. Some never should.
Boring, inspectable pipelines. The agent systems that survive are embarrassingly simple: fetch → process → draft → log → notify. Every step leaves a trace. When something goes weird — and something always goes weird — you read the log and know why in minutes. Clever autonomous architectures demo better and die faster.
Agents watching, humans deciding. Monitoring is the most underrated agent job: watch the metrics, the competitors, the mentions, the error logs; summarize what changed; escalate what matters. It converts “we should keep an eye on that” — which no one ever does — into a system that actually does.
What fails
Unattended output to the outside world. Every horror story shares one root cause: machine output reached a customer, a public page or a codebase without a human in the loop. The fix isn’t better prompts. It’s the gate.
Agents pretending to be senior. An agent will confidently produce a strategy, an architecture, a legal opinion. It reads as senior. It is not senior. It’s the world’s best-read junior — brilliant at execution inside a frame, dangerous at choosing the frame. Companies that skip the human senior don’t save money; they defer the invoice.
Automating a broken process. If your reporting is chaos, an agent produces automated chaos, faster. We always fix the process first — usually simplifying it — and only then automate. Half the value of an “AI transformation” is the cleanup it forces.
The rules we won’t break
- Every agent action is logged. No exceptions. Trust comes from audit trails, not vibes.
- External-facing output gets human sign-off until a system has earned specific, bounded autonomy.
- Client data stays in enterprise APIs or client infrastructure. Never in tools that train on it.
- The off switch is one command. Any system we build can be paused instantly by the client, without us.
Where to start
Not with a moonshot. With the workflow everyone on your team complains about: the weekly report, the inbox triage, the content adaptation, the data cleanup. Small blast radius, obvious value, measurable time saved. Get one boring win running, and the second system builds itself politically.
That first system typically takes us days to build, not months. Which is the other thing the disillusioned literature gets wrong: the cost of trying is now low enough that the analysis paralysis costs more than the experiment.
This is what our AI & automation service does. The audit that finds your first three systems is a fixed-price sprint.