Meta says a production AI system became more reliable when it stopped asking one agent to know everything. Its thin orchestrator routes work to specialised agents, and Meta says the change cut preparation for a new data column to under an hour, from three to four days with its earlier system.
Meta AI Watch analysis
What happened
In an account published by Meta Engineering, the company describes automating data-column onboarding across three code repositories, three programming languages and as many as six pipeline stages. Its first attempt used one agent loaded with domain knowledge. Meta says that system began mixing up concepts across subsystems and produced unreliable results in production.
The replacement has a roughly 1,300-line orchestrator that routes more than 20 workflows to 14 specialised agents. The orchestrator handles sequencing, required information and reliability checks, while the specialist agents do the domain reasoning with separate, focused context. Meta says about two-thirds of the system’s changes land without significant human edits.
Why it matters
The useful lesson is architectural, not mystical: adding more context to one agent can make a complicated task worse. Separating predictable routing from specialised reasoning gives each component a clearer job, and makes it easier to contain a failure or reuse existing agents in a new workflow.
Meta’s figures describe its own system, not a controlled comparison proving the same design will work everywhere. But the concrete time reduction, alongside the account of what failed, makes this more useful than another abstract promise that agents are ready for production.
Our read
“Make the agent smarter” is a tempting default. Meta’s experience suggests that, for multi-stage work, making the central agent do less may be the more practical upgrade. Teams building agent systems should look closely at where routing ends and domain reasoning begins, rather than handing one model the whole filing cabinet and hoping it develops a good sense of shelving.
What to watch
- Whether Meta reports similar results as it adds pipeline domains and workflows.
- How much human review is still needed beyond the changes it says land without significant edits.
- Whether other teams can reproduce the benefits of isolated specialists and deterministic routing.
Discussion spark: For production AI agents, is the safer design a deliberately simple router with specialist workers, or does splitting work across agents create too many hand-offs to manage?
Sources and evidence
- The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do (19 October 2026, 15:30 UTC)
not affiliated with, endorsed by, or operated by Meta