$85,000 in tokens later: What I learned from scaling agentic coding at Lovable

agentic-codingai-code-reviewdeveloper-workflowsubagentspull-requestsautomation

Abstraction: Scaling agentic coding workflow lessons at Lovable

Key points:

Author's token spend went from ~$600/mo pre-Lovable to ~$25K/mo in May, ~$85K total since January; productivity rose from 20-30 merged PRs/week to 150+; merged 293 PRs in first week of June with no traced production defects.
~75% of tokens go to implementation, ~25% (growing) to automation/AI reviews; one human now oversees 6-7 agents, each with its own subagent swarm.
Human review is reserved for high-impact decisions via RFCs/ADRs, not line-by-line; a markdown-policy-file AI classifier routes PRs into fast-AI, slow-AI, or human review lanes by risk (infra/auth = always high risk).
Advocates small stacked PRs (tens-hundreds of lines) — AI review quality drops sharply on big PRs; a 6K-line PR passed AI review but splitting it surfaced real issues.
Tooling/practices: Beads as external task tracker for >1M-token tasks, subagents for delegation, Graphite "merge when ready" wrapped in unit-tested scripts, /clear per task, agents write their own skills; "tokenmaxxing" is foolish — outcomes matter, ~5% attention to efficiency.

Connections: Lovable · Graphite · Beads · Agentic Coding · AI Code Review · AI Agents · Context Management