Loading…
Free-text scratchpads. Unbounded tool use. Recursive critics. We've made every mistake. Here's the postmortem.
Multi-agent systems fail in ways single agents do not, and most of those failures are architectural rather than model-related. Five we have run into.
Give two agents a conversation and no termination condition and they will be endlessly agreeable, each deferring to the other, burning tokens on consensus nobody needed. Every loop needs a stop rule that does not depend on an agent deciding it is done.
A coordinator's job is routing and arbitration. The moment it starts doing the task itself, the specialists become decoration and you have a single agent with extra latency.
A scratchpad every agent can write to becomes a scratchpad every agent can corrupt. Give each field an owner, and make the others read-only against it.
Three agents that each retry three times is not three retries. It is up to twenty-seven, against a rate limit sized for one. Budget retries at the system level, not per agent.
When a single agent misbehaves you read the prompt. When five do, you need a trace showing who called whom with what, or you are guessing. Build the trace before the fifth agent, not after.
If you cannot draw the call graph on a whiteboard, you cannot debug it at two in the morning.
Why the next decade of competitive advantage belongs to organizations whose software does the doing — and what that demands of how we build.
A practical guide to building eval suites that actually catch regressions — and why your prompt is code that needs CI.
Hybrid search, reranking, citations, permissions, drift monitoring. A no-nonsense reference architecture.
One call with a Kriyava AI architect. We map your highest-leverage workflow, scope a build, and ship something live before your next quarterly review.