Here's a thing I can do that sounds like a party trick. I open a brand-new session with my agent and type, almost word for word: "take ticket 123 and do it." And it does — the whole thing, end to end, without me hovering. Not because I found a magic prompt. Because I spent months teaching it what "as usual" means.
I bring this up because of a complaint I keep hearing, some version of: "I can't get my agent to just do a task. I set it loose and then I have to walk behind it fixing everything." And look — if you started last week, that's completely fair. I couldn't either at first. But when someone's been at it for months and still can't get the agent to carry a task cleanly, I get suspicious. Not of the agent. Of how they're working with it.
"As usual" is a lot of things
Let me show you what actually happens when I say "do it as usual," because the phrase is hiding an entire workflow the agent has internalized:
- It takes the ticket and assigns it to me.
- It moves it to In Progress.
- If the ticket was sitting in the backlog, it pulls it into the current sprint.
- If it has no story points, it estimates it — and if it genuinely can't tell, it asks me, the way we agreed it would.
- It creates a branch by our naming convention, after first updating main so it's not branching off something stale.
- It makes the changes, committing as it goes with human-readable messages that say what was actually done.
- It opens the pull request, writes a real description, adds the necessary context and links, mentions the people who need to see it, and links the PR back to the ticket.
- Then it waits for the pipelines, checks whether anything went red, and fixes it if it did.
All of that from four words. And none of it was true on day one.
The part nobody wants to hear
That list didn't arrive as a single clever configuration. It arrived one failure at a time. The agent would branch off stale main — so I taught it to update first. It would write commit messages like "fix stuff" — so we worked out what a human-readable message looks like. It would guess a story-point estimate confidently when it had no business guessing — so we agreed that when it's unsure, it stops and asks me instead of inventing a number.
Every step in "as usual" is a scar. It's there because something once went wrong, and I turned the fix into a rule instead of re-explaining it every session.
That's the whole method, honestly. It's not prompt cleverness. It's accumulation. Each time the agent got something wrong, I didn't just correct it for that one run — I wrote the correction down somewhere durable, so the next session started from the corrected baseline. Do that for a few months across a dozen steps, and "as usual" stops being a phrase and becomes a reflex the agent actually has.
Give it real context, not guesses
The second half of this is tooling. An agent working from guesses is an agent you'll always be correcting. So I connected the things it needs to know the state of the world instead of hallucinating it: an MCP server to read Jira, another for Confluence, one for our docs, one for Postman — the tools the work actually runs on.
The difference is night and day. When the agent can read the ticket, it doesn't misunderstand the task. When it can see the docs and the API collection, it doesn't invent an endpoint. The tools aren't there to make the agent look impressive — they're there so its decisions are grounded in what's real. This is the augmented-LLM idea done properly: one agent, made capable by good tools, is worth more than a crowd of agents guessing at each other's intent.
One agent, not one per repo
I told this story in the multi-agents piece, but it belongs here too, because it's the same instinct. My project spans six or seven repositories. My first move was to set up an agent in each one. Then I noticed my tasks almost never span repos — they live in one, and multi-repo changes have come up maybe twice. So instead of maintaining seven configurations, I put every repo in one workspace and taught one agent. Every hour I didn't spend duplicating setup across seven repos, I spent making the single one better. That's why it can do "as usual" at all — the investment was concentrated, not scattered.
The decision points matter most
The subtle part — the thing that separates an agent that helps from one you babysit — is deciding where it should act and where it should stop and ask. I didn't tell it to just do everything autonomously. We agreed on the seams. Estimation it can't confidently do? Ask. Convention it knows cold? Just do it. That negotiated boundary is what makes "as usual" safe to run unattended. The agent isn't guessing about my intent at the risky moments, because we already talked through those moments and it knows the rule.
This is why "just be more autonomous" is bad advice. Autonomy isn't a slider you crank to the top. It's a map of which decisions the agent owns and which ones come back to you — and you build that map the same way you build everything else here, one correction at a time.
The reframe
So when I hear "I can't get my agent to do a task in one shot," what I mostly hear is a missing investment. People want the "do it as usual" outcome without the months of turning mistakes into durable rules and wiring in the tools that ground the agent in reality. The agent that runs my whole ticket workflow from four words isn't smarter than yours. It's just been taught more, and the teaching stuck because I wrote it down instead of repeating it.
The uncomfortable truth is that this looks a lot less like prompting and a lot more like onboarding a junior engineer who happens to have perfect recall. You explain the convention once, correctly, and it never forgets it — but only if you actually captured it instead of hand-waving it in a single session that evaporates when the context does.
Pick the one task you do most often with your agent and write down every correction you find yourself repeating this week. That list is the start of your own "as usual." Send it to me if you want a second pair of eyes — I'll tell you which corrections are rules and which are you doing the agent's thinking for it.
Top comments (0)