What three windows of our agents’ sessions taught us about handover points and token cost

This is the second article on how our agents use tokens. The first one counted a week of 65,000 requests and ended with a hypothesis: if the size of the context multiplies the weight of every step, keep it short. We then changed several settings and measured again, twice: a day after and a week after. The two measurements showed different things, and the difference is what taught us most.

What we changed

On 29 September we switched on four things:

• Handover at 40% of the window: when an agent’s context reaches it, harnsy asks the lead to hand the role over to a fresh agent, which starts from a written note. A threshold set on a role overrides the default.
• Sonnet 5.5 on six roles (at the time) that follow a procedure or write text. Developers, reviewers and leads stayed on Opus 5.5.
• Shorter role documents: the developer’s and the lead’s came down to about a third of their length.
• Batched notices: 299 batches instead of 1,086 single notices in the first day.

Our own model of the step history said this would take about 20% off the weight of the main sessions. In the first 17 hours the step price fell by about a quarter, which agreed.

How we compared the three windows

Same method as in the first article: the transcripts Claude Code keeps, one line per request, every request priced at its model’s published API rate, no duplicates. We limited the data to our company’s sessions and recomputed the “before” window the old way, over all our projects: it gives the same figures as the first study (73,046 steps, $7,044), so the comparison is fair. Because the data is narrower, the average context before is 305k here, against 306k over all projects.

A day later: a step cost a quarter less

The step of a main session, the one an agent makes in its own conversation, cost $0.096 before and $0.075 right after. The average context of a step fell from 322k to 257k tokens. The weight in steps above 400k tokens fell from 48% to 14%.

The reason was the handover point. A step on the same context costs the same before and after ($0.089 and $0.090 for 200–400k); the saving came only from contexts growing less often. Our estimate of what each change was worth:

A week later: the handover point decides

We measured again on 5–8 October, three full days. The step of a main session now costs $0.101 (it was $0.096 before our changes and $0.075 the day after). The average context is 311k tokens, close to the 322k we started from. The weight in steps above 400k is 36%. What differs from the day after is the handover points, as the next section shows.

All steps together, subagents included, cost $0.087, against $0.091 before. That figure blends two things: subagents now carry 20% of the weight (3% before), and their steps are cheap, $0.055. The main sessions, where the settings show, are the better yardstick.

What moved it: the handover points

The cause is a setting. In the first days of October the handover points were raised: developers and reviewers to 600k, the site lead from 600k to 700k; the product lead had been at 700k since 30 September. The numbers below show where sessions actually handed over.

The roles that kept the 40% setting (designer, DevOps, UX) stayed near their numbers of the day after. The roles with higher points carry longer contexts and weigh more per step.

A higher point is a trade-off, and we measured one side of it. A successor that starts from a note can lose the thread; a higher point lowers that risk. We measured what the higher point costs in weight; what it protects, we did not measure.

The expensive roles kept their models. Sonnet’s share of the weight fell from 27% to 11.5% only because the roles on Sonnet worked less in these three days.

Waiting has a price

The second lever is waiting. An agent that sits idle for more than an hour loses its cache and has to write the whole context again. Before our changes this happened 10 times a day; now 65 times, at $2.04 each, 14% of the weight.

It is not only the morning. Re-warms run all day, with peaks at 10:00 and around midnight. The median pause is 2.7 hours; a quarter of the pauses are shorter than 1.5 hours and a quarter longer than 8. Leads and developers are the most frequent, then DevOps, UX, reviewers and the designer. An agent waits for a review or for the next task, and the cache expires in the meantime.

Messages wake the agents with the largest contexts

Turns started by a message from another agent are now 65% of the weight (69% with subagents’ hand-backs), against 79% before. The drop is mostly the subagents taking part of the work, not a change in how agents talk.

Short turns, one to three steps like “got it”, are 729 a day and 17% of the weight, $0.225 per turn. Leads make 64% of them. Per 1,000 steps they went from 101 to 74 to 65, a slow decline. Each such turn costs more than in the first study ($0.225 against $0.158), because the leads’ contexts are larger at their higher handover point.

Subagents took work and stayed cheap

Subagents went from 3% to 20% of the weight. Almost all of them are developers handing a whole build to a subagent. Such a subagent does not hand over and grows without a threshold: the largest reached 783k tokens and weighed $61. Even so, steps above 400k are only 2.9% of the subagents’ weight, and their step is cheap. Moving work into a subagent is useful: the main session grows more slowly.

What we would try next

These are proposals, not results. Each estimate is a replay of our past steps, with the handover price set at $0.70 and a fresh session starting at 60k tokens. A real lead’s handover note is longer, so the saving for leads may be a little smaller. The percentages are shares of the whole weight, about $964 a day at API list prices.

• Lower the points: developers and reviewers to 400k, leads to 500k, for about −11%. If the higher points protect the thread of the work, the compromise is 500k for all three, about −9%. Run it on one team for a week and compare not the cost of a step but the cost of a closed task.
• Keep the cache warm for an agent that waits: every 55 minutes of idling, for up to four hours, harnsy wakes it with an empty short ping. Reading a 322k context from cache costs about $0.06, re-warming about $2. By our window that is −6…−7% (a ping budget of $14 a day included). It needs a feature in harnsy, and we do not know how Anthropic counts pings against a subscription limit.
• A digest for the lead instead of separate “for your information” messages: acknowledgements pile up and arrive in one batch every 10–15 minutes; questions and blockers go at once. If half the leads’ short turns disappear, that is about −5%. Not built yet.

What we would not do: limit subagents to save money, or widen Sonnet only for the price; its step costs almost the same ($0.074 against $0.089 for Opus 5.5 in this window).

What we measured and what we did not

What to take from this

• What a setting saves lasts as long as the setting. Measure again after a week, on a full window.
• Compare the price of a step and the shares, not daily totals: the load changes.
• Context size at handover is the biggest lever we found. The higher the point, the more every following step weighs.
• Waiting has a price too: a cache that expires is written again.
• Before you move a handover point, decide how you will check the quality of the work, and look at it next to the cost.

Where harnsy fits

harnsy is a local app that connects Claude Code, Codex and OpenCode agents into one team with roles. The handover point is a setting per role: the default is 40% of the window, and the dashboard shows where each role stands. A higher point is a choice between money and continuity, and harnsy does not make it for you. How it installs, and what it changes on your machine, is on the install page.