The Hardest Part of an Agent Isn't the Brain

Nearly half of teams say the top barrier to agents is integration, not intelligence. The brain was the easy part. The plumbing is the job.

Ask teams what is blocking their agent projects and the answer keeps coming back the same: it is not the model. Nearly half say the number one barrier is integration with existing systems. Not smarter reasoning. Not bigger context. Getting the thing plugged into the business without breaking the business.

Anyone who has built one of these is nodding. The demo is the model being clever. The product is the model being allowed to touch your systems safely, and that second part is where projects go to stall.

why integration eats everything

The model is the easy part because it is the standardized part. You call an API, you get intelligence back. Done. Everything after that is custom.

Permissions. Who is the agent acting as, what can it see, what can it change, and who approved that? Most companies cannot answer those questions cleanly for their human employees, let alone for software acting on their behalf. Minimum necessary access is a nice principle until you try to implement it across fifteen systems that were purchased in four different decades.

Then the long tail of "but in our case." The ticket queue with the custom fields. The database with the schema from 2009 that nobody is allowed to touch. The approval workflow that lives in someone's head. Every business has a layer of undocumented reality, and agents trip on exactly that layer, because the demo never includes it.

Then maintenance. The model updates, the API changes shape, the business renames the regions. An integration is not a thing you build once. It is a thing you operate, and most agent plans budget for the build and forget the operate.

what the winners do

The teams getting agents into production share a playbook, and it is boring in the best way.

Read-only first. Let the agent see everything before it touches anything. Reading is cheap and safe. Writes are where the risk lives. The agent earns write access one workflow at a time, with a human checkpoint on anything that cannot be undone.

Narrow the surface. One system, one job, one team. The integrations that survive are small enough to understand completely. The ones that die tried to boil the ocean on day one.

Log everything. Every read, every proposed write, every override. Not because you do not trust the agent, but because when something goes wrong at 2am, "the agent did it" is not an incident report. Auditability is what turns a demo into infrastructure.

Treat it like software. Version it, test it, monitor it, page someone when it breaks. The teams that survive are the ones that stopped treating agents as magic and started treating them as systems. Unsexy. Effective.

the honest read

The industry spent two years obsessing over the brain and is now discovering the body. This is the normal order of things. The exciting part arrives first, the hard part arrives second, and the hard part is where the value lives.

If you are planning an agent project, budget accordingly. The model is maybe twenty percent of the work. The other eighty is permissions, plumbing, logging, and the long patient business of connecting new intelligence to old systems without breaking either. That is the job. It was always the job.