Why are agents harder to use outside coding?

Coding agents get a lot of help before you even write the prompt. They start with a codebase that shows how the system works and how similar problems have been solved. The place where they do the task also holds much of the context they need.

They can then run the code and test their changes. Tests do not catch every problem, but a failed test tells the agent to keep working. It can fix the issue and try again without asking you to check each step.

A finished file can still be bad work

Now ask an agent to make a customer presentation. It may need something the customer told you last week, or the reason your team chose this plan. Some of that is in documents. Some is only in your head. The agent can make the slides, but it lacks the knowledge that shapes the story.

The slides can open and look fine while making a poor argument. The agent calls the task done, but the work still falls short of your team’s standards. You become the reviewer, sending it back for changes and checking each new draft.

Build the support for autonomy

Review gates let the agent get that feedback before it brings you the work. They check its draft against the standards you set and tell it what to fix. Each new version faces the same checks, so fixing one issue does not let another slip through.

The team also needs to keep learning from the work. When you do a task yourself, you learn why certain choices work. When an agent does it, the finished file does not give you all that understanding. Useful lessons need to carry into the next task, so you do not have to explain the same things again.

We built Firmament to bring that context and review together. Its context layer keeps useful knowledge from the work agents do with your team. Its review gates give them an automated check against your standards, so they can revise their work before handing it back to you.

That is how I think about autonomy: how much work can you give an agent without having to guide every step? Coding gives it a head start. Other teams need to build that support into how they work.