TECHNICAL MEMORANDUMTM-26-005
SUMMARY. For seven weeks, one developer (me) paired with an AI coding agent to build a product: 54 delivered work items in the three most intensive weeks, at roughly USD 36 each. The surprise is where the money went: 65% of spend was the agent re-reading its own conversation history. Session hygiene, not model choice, is the cost lever.
1. BACKGROUND. Three of us are building a bootstrapped product on the side: five repositories, first commit to production deployment in seven weeks, most of the implementation done by an AI coding agent. The agent's tooling records every API request, so I ran the numbers for my own pairing with it: token-level costs for 69 working sessions, converted to API list prices. I then cross-checked the pattern on two smaller side projects of very different shape - a hardware hobby project and a documentation project.
2. FINDINGS.
2.1 Output first: 635 commits across the five repositories. In the three-week window with full token data, 54 delivered work items; in the deployment repository the agent co-authored 97% of commits.
2.2 The bill for those three weeks, at list prices: about USD 1,921. That is USD 36 per delivered work item, USD 3 per commit, USD 1.64 per prompt I typed.
2.3 Where it went: 65% of spend was cache reads - the agent re-reading the session's accumulated context on every request. Freshly typed input was near 0%, the model's own output 13%. Both side projects showed the same split, with cache reads at 60–65%.
2.4 The expensive pattern is the immortal session: kept open for days as a workspace. Twelve such sessions consumed 59% of all spend. The worst began as one small question and ran three days, 133 prompts, USD 400. A marginal prompt in a mature session costs USD 2–3; in a fresh one, about USD 0.50 - however simple the question.
2.5 The cheap patterns: sessions scoped to a single task averaged USD 0.56 per prompt, four times cheaper than the long-lived ones. All exploration delegated to subagents together cost 2% of total.
3. ASSUMPTIONS. All dollar figures are API list-price equivalents computed from recorded token usage; the account runs on a flat subscription, which bills differently. Read them as a resource- intensity metric: the relative numbers are the lesson. One developer's practice, three projects, consistent direction.
4. CONCLUSIONS. Cost is context size times request count; everything else is detail. The fix is not a cheaper model but a working discipline: one session per task, and a fresh session when the topic changes. That is only cheap if the durable state lives outside the conversation - in plan documents, tickets, and decision ledgers a new session can re-read in seconds. Documentation discipline turns out to be the actual cost optimization. Each model generation raises the price of bad session habits.
5. ACTION. The working rules this yields: new topic, new session. Treat context compaction as a smell that one session should have been two. Delegate exploration to subagents. Script anything the agent does more than a few times. Keep decisions in files, not in the conversation.