I Spent $847 Running AI Agents for a Month. Here's the Spreadsheet.

Last month I went all-in on AI agents. Not just copilot-style autocomplete — full autonomous agents that browse the web, write code, manage deployments, and chain together multi-step workflows.The bill was $847. Here's what I learned about where the money actually goes.## The SetupI had three main agent workflows:1. Code review agent — reviews PRs, leaves comments, suggests fixes2. Content pipeline agent — scans trends, writes drafts, schedules posts 3. Monitoring agent — watches dashboards, sends alerts, auto-remediates simple issuesEach one runs on a different model. Each one has different token economics. And that's where things get interesting.## Where the Money GoesHere's the breakdown that surprised me:| Agent | Monthly Cost | Tokens Used | Cost/Task ||-------|-------------|-------------|-----------|| Code Review | $312 | ~18M tokens | $2.60/PR || Content Pipeline | $389 | ~24M tokens | $4.85/article || Monitoring | $146 | ~8M tokens | $0.12/check |The content pipeline was the most expensive — not because it used the most tokens per task, but because each article requires research (web browsing = lots of tokens), drafting (long outputs), and revision cycles (multiple round trips).Monitoring was the cheapest because most checks are "everything's fine" — short input, short output, done.## The Hidden Cost: Agent Reasoning DepthHere's what nobody tells you about agent costs: the expense scales with reasoning complexity, not usage volume.A simple monitoring check: 200 tokens in, 50 tokens out. Done.An agent that needs to understand a complex PR, check for security issues, cross-reference with coding standards, and write a nuanced review? 8,000 tokens in, 2,000 tokens out. Per file.I found that 80% of my costs came from 20% of tasks — the ones where the agent had to actually think.## What I Did to Cut Costs by 60%### 1. Model RoutingNot every task needs the smartest model. I set up a router:- Simple checks → small/fast model ($0.10/1M tokens)- Standard tasks → mid-tier model ($3/1M tokens) - Complex reasoning → top-tier model ($15/1M tokens)This alone cut costs by 40%.### 2. Context Window ManagementAgents love to accumulate context. A conversation that starts at 500 tokens balloons to 50,000 tokens after 20 turns — and you're paying for every token in every API call.I started spawning fresh sub-agents for each independent task instead of keeping one long-running agent. Each sub-agent gets only the context it needs.### 3. Caching Repeated PatternsMy monitoring agent was re-analyzing the same dashboard layout every 5 minutes. I cached the layout analysis and only re-ran it when the page structure changed. Saved ~$90/month.### 4. Kill Unnecessary ReasoningThe biggest win: I stopped asking agents to "explain their reasoning" for routine tasks. A monitoring check doesn't need a 500-word explanation of why everything is fine. "OK" is enough.## The SpreadsheetHere's the tracking template I use:| Date | Agent | Task | Model | Input Tokens | Output Tokens | Cost | Success? ||------|-------|------|-------|-------------|---------------|------|----------|| 2/1 | Review | PR #142 | opus | 12,400 | 3,200 | $0.47 | ✅ || 2/1 | Monitor | Health check | haiku | 180 | 45 | $0.00 | ✅ || 2/1 | Content | Draft article | sonnet | 8,900 | 4,100 | $0.08 | ✅ |Track everything for the first month. You'll be shocked at what's actually expensive.## Key Takeaways1. Agent costs scale with complexity, not volume. 10,000 simple tasks can cost less than 100 complex ones.2. Model routing is the single biggest cost lever. Use the right model for the right task.3. Fresh contexts beat long conversations. Spawn sub-agents instead of maintaining marathon sessions.4. Measure before optimizing. Track every API call for a month before making changes.The $847 was honestly worth it — the agents saved me about 60 hours of work. At my billing rate, that's a huge win. But without tracking, I'd have no idea if I was getting good ROI.If you're running AI agents in production, track your costs obsessively. The difference between a well-optimized agent setup and a naive one can be 5-10x.---I write about practical AI tools and developer productivity. If you're exploring AI coding assistants, I put together a collection of 50 tested prompts for common dev workflows — the same ones I use to keep my agent costs predictable.