Compound Engineering for Solo Developers: 80% Non-Coding Time and What It Costs in Tokens

The Method

Every, the software media company, has open-sourced its "compound engineering" method: one engineer maintains five products, spends only 20% of their time writing code, and 80% on Plan → Work → Review → Compound loops. The magic word is the fourth step — every problem solved gets encoded into CLAUDE.md and docs/solutions/, so the next time the AI hits a similar issue, it avoids the pit automatically.

Their open-source plugin ships 26 specialized agents, 23 workflow commands, and 13 skills — zero-configuration. Two of the standout commands: /workflows:review spawns 14 parallel agents to review code, and /workflows:plan in ultrathink mode fans out 40+ research agents.

Why This Is Interesting for Solo Devs

A solo developer maintaining even one product is stretched thin. The idea that one person could maintain five with disciplined AI orchestration is either revolutionary or a very expensive way to fail. The cost of the "or" is what most solo devs need to see before they try it.

Baseline: One Product, Traditional AI Coding

Typical solo developer using Claude Code or Cursor on one product:

• 5 hours/day of AI-assisted coding
• ~150K–300K tokens/day input, ~30K output
• At Claude Sonnet 5 promo pricing ($2/$10): ~$0.90–$1.20/day
• Monthly: $25–$40, comparable to a Claude Pro subscription

Compound Engineering: Five Products, Same Solo Dev

The compound method fundamentally changes the token profile because so much time is spent orchestrating parallel agents:

• 5 products × 1 hour/day/product of orchestration = 5 hours/day
• /workflows:review spawns ~14 agents per invocation; if used 3x/day = 42 agent invocations
• /workflows:plan in ultrathink mode: 40+ agents per planning session; assume 1/day = 40 agents
• Regular coding calls: ~2–3 hours across the five products

Rough token math:

Monthly bill: ~$570. That's a 15–20x jump from a solo developer running one product on Claude Pro.

Does It Pay Off?

Compare $570/month against the alternative of not maintaining five products. If the five products together produce $3K/month in revenue (a reasonable indie SaaS target), the AI bill is 19% of gross. That's tolerable — comparable to what a small SaaS pays for cloud hosting.

If they produce $10K/month, the ratio drops to 5.7% — quite good. If they produce $500/month, the ratio is 114% and the model is broken.

Cost Reduction Levers

The compound method's raw cost is high, but it has more optimization slack than a solo dev on one product does:

1. Route reviewer agents to cheap models. Sonnet 5 for regular coding, DeepSeek V4-Flash ($0.14/M input) for the 14 reviewer agents. Cuts review cost 90%+.
2. Aggressive prompt caching. The CLAUDE.md-style context that gets loaded in every agent invocation is exactly what prompt caching was designed for. Saves 60–80% on repeated context tokens.
3. Async batching for planning. Planning workflows don't need instant response. Batch APIs are 50% cheaper.
4. Smaller planning fanout. Do you need 40 research agents, or would 10 suffice? Cut to what actually informs the plan.

Realistically, all four together can cut the compound-engineering bill to $150–$250/month, which makes it viable even for products at the $1K/month revenue mark.

Where Compound Engineering Wins

• You're a technically strong solo developer who could afford a $500/month AI bill
• You already have several products (or plan to launch several) that share code
• You're disciplined about documentation — the "Compound" step only works if you actually update CLAUDE.md
• Your revenue depends more on breadth of products than on depth of any single one

Where It Fails

• You're just starting out and don't have revenue to justify $200+/month AI spend
• Your product needs deep, novel technical work — parallel agents can't replace expert focus
• You're the kind of developer who resents writing docs — the compound loop only compounds if you keep the notes current
• You spread yourself so thin that no product gets the depth it needs

The Realistic Starting Point

Don't jump to five products. The compound method's real value shows up gradually:

1. Adopt the CLAUDE.md discipline on your current project. Update it every time you solve a non-obvious problem.
2. Add one /workflows:review-style command that fans out 3–5 reviewer agents. Measure the quality delta.
3. Add a second product to the mix only after the first one is running smoothly with compound discipline.
4. Track cost per delivered feature, not per token. If cost per feature is dropping while feature velocity holds, keep going.

Bottom Line

Compound engineering is real and works, but the token bill is 15–20x the naive solo dev baseline unless you optimize. With aggressive routing, caching, and batching, the bill drops to 4–8x baseline — which is worth it if the products it maintains generate real revenue. For hobbyist developers, this is currently too expensive. For indie founders shipping revenue-generating products, it's a serious competitive advantage.