They replaced one do-everything agent with 31 narrow bots — 6 lessons from a year of production pain

tasqbots · reddit · 2026-09-08

A team building SMB automation (invoice reconciliation, HR onboarding, document extraction, support triage) shares lessons after their general agent fell apart in production and was split into 31 narrow bots.

Narrow scope makes evals tractable: 8 failure modes per bot vs unbounded surface for a general agent.
Users don't want a chat box: free-text prompts hurt usage; they want a "Reconcile March" button.
The model is the cheap part: 80% of effort goes to retries, idempotency keys, audit logs, human approval.
Never write directly to the system of record: a bot was 97% accurate writing to client books; the 3% cost two days of reversals. Now bots propose, humans approve, and the approval writes.
Log the input, not just output — you'll need the exact document weeks later.
Narrow bots sell better: "reads vendor invoices and flags mismatches" closes; "AI automation platform" doesn't.

Self-critique: 31 bots is a lot of surface to maintain; if starting over, they'd build the approval/audit/retry layer first and treat each bot as config.

Original post →

They replaced one do-everything agent with 31 narrow bots — 6 lessons from a year of production pain

More from coding & agent