vibencode
Agents and orchestration
Can the whole process run without someone driving it?
“Let’s get agents doing this end to end.”
Agents with named roles and scoped tool access, state that survives a crash, a retry path that does not loop forever, and a human approval before anything writes.
Most of what makes an agent system work is not the agents. It is the plumbing around them: where the state lives so a crash resumes instead of starting over, what happens when a call times out halfway through a sequence, which step is allowed to write to a customer without a person looking first. A demo gets to skip all of it. That is why demos take a fortnight and production takes a quarter, and why the list below is mostly plumbing.
- 01Name the roles, and give each the smallest tool access that still does the job.
- 02Design the state: what has to survive a crash, what has to be safe to run twice.
- 03Write the failure paths: retries that stop, fallbacks that degrade, and a dead end that pages a human.
- 04Put approval exactly where money, data or reputation changes hands, and nowhere else.
- 05Evaluate the whole run rather than the model, because that is what fails.
- Agent architectureThe roles, what each one is allowed to touch, and where the run may continue without you.
- Tool integrationsYour real systems, behind your access controls, with credentials scoped per role rather than shared.
- Approval & recovery pathsWhere a person says yes, what happens when they are not there, and how a half-finished run is picked up.
Autonomy is not a setting. It is a series of decisions about what an error costs.
Map and cost
The process as performed, and what a mistake costs at each step. The second half is the one people skip.
Build the run
Roles, scoped tools, durable state, retries with ceilings, fallbacks that degrade rather than stop.
Shadow
Real volume, no writes. We watch where it stops, and turn each stop into a handled path.
Live, then hand over
One team or region first, with the pager pointing somewhere real and a runbook that names the failure modes.
That is the shape of a full engagement. If it is more than you want to commit to yet, the same people will spend two days on one piece of it first, and you are free to stop there. Start a 48-hour sprint
- Credentials for the real systemsSandbox access hides the rate limits, the timeouts and the record that has been broken since 2019, which is the run that will fail.
- The person who runs it todayA few hours a week. They know which of the eleven steps is the one that actually goes wrong, and no document has ever contained that.
- A decision on what runs unattendedWhere a mistake is recoverable and where it is not. We bring the question with the costs attached; someone senior has to answer it.
- Somewhere for the system to pageAn on-call route that exists and that someone answers. An agent failing silently is worse than the manual process it replaced.
What actually is agent orchestration?
Several AI agents, each with a named role and only the tool access that role needs, running a multi-step process end to end. The orchestration is the part that decides which agent runs when, what holds the state between them, what happens when one fails, and where a person has to say yes. The agents are the easy half.
How autonomous can an agent system actually be?
More than most teams expect for retrieval and assembly; less than most vendors imply for anything that writes. The honest answer is per step rather than per system, and working it out is most of week one.
Which framework do you use?
Whichever survives your constraints. We have shipped on several and have no loyalty to any of them. The architecture of roles, state and failure paths outlasts the framework, and is the part worth arguing about.
Can we start with one step rather than the whole process?
Usually the right move. One step running unattended, with the pager wired up, teaches you more in a fortnight than a full pipeline sitting in shadow mode for a quarter.
What happens when a model provider changes something?
The run is evaluated against a fixed set, so a regression arrives as a number rather than as a support ticket. Changing a model becomes a planned change with evidence attached.
We already have a prototype that half works.
Bring it. Reading what broke is faster than starting again, and the place it breaks is almost always the state handling rather than the model.
If you only need answers out of your own documents, an agent system is heavier than you need. A grounded assistant is cheaper to build, cheaper to run and far easier to be confident about. That is GenAI implementation.
Can the whole process run without someone driving it?
Two days on one step of the process. Written up at the end.