All services
Practice area
Agent architecture & orchestration
Agent projects tend to stall in the same place. The demo goes well, everyone gets excited, and then production finds the seams. The loop won't end. The specialists duplicate each other's work. Something goes out the door that your own policy says shouldn't have. None of that is a prompt problem, and none of it gets better by writing a longer prompt.
You probably need this if
- Your agent either runs forever or quits early, and nobody can predict which.
- A hard turn limit is the only thing keeping the bill down.
- Split the work across several agents and they either duplicate each other or leave gaps.
- A rule you can't afford to break is written in the prompt, which means it holds most of the time.
- When it hands off to a person, they get the conversation and none of the context.
What you get
- An agent that knows when it's finished
- Most loops stop for the wrong reason: a turn cap, or text that happened to sound like a conclusion. Both are guesses, and both fail in the direction that costs money. We tie termination to the actual signal, and make sure each pass can see what the last one learned.
- A coordinator worth having
- One agent owns the plan. It works out what the job needs, calls only the specialists that job requires instead of marching everything through the full pipeline, and stays in the middle of every exchange — so when something goes wrong there is one place to look.
- Handoffs that carry their context
- A specialist agent starts with nothing. Whatever the last one found has to be handed over on purpose, with sources kept separate from findings so attribution survives the trip. A surprising share of "the agent made it up" turns out to be this.
- Rules that hold every time
- Anything you'd be uncomfortable explaining to a regulator does not belong in a prompt. Those checks go in code that runs before the action does, so the limit is enforced rather than requested.
- Work you can pick back up
- Long investigations shouldn't start from zero every morning. Named sessions to resume a thread, branching to compare two approaches against the same groundwork, and a clear rule for when yesterday's findings have gone stale enough that starting clean beats carrying on.
The judgment behind it
These are the calls that decide whether a system survives contact with production or just demos well. They are not obvious, and getting them wrong is rarely obvious either — until it is.
- Prompts ask. Code enforces.
- A prompt is a request with a good success rate. That's fine for tone and unacceptable for confirming who someone is before you move their money. If a rule has to hold every single time, it belongs somewhere the model cannot talk its way past.
- Fan out in one move, not three.
- Asking for five things at once and asking for five things in a row look nearly identical in a transcript and cost wildly different amounts of time. A good number of "the agent is slow" complaints are really this.
- Say what good looks like, not what to do.
- Step-by-step instructions cap a specialist at whatever its author thought to anticipate. Give it the goal and the standard instead, and it can handle the case nobody saw coming.
- Fixed pipeline or adaptive plan — choose on purpose.
- Predictable work with known steps wants a fixed chain. Open-ended investigation wants a plan that rewrites itself as it learns. Picking the wrong one is why some agents feel rigid and others feel aimless.
Engagement shape
Usually 3–6 weeks: architecture review, a working reference implementation, your business rules moved into enforced checks, and a written handoff your team can build on.