SH Shamim Hossain
Going further · about 5 minutes

The rereading tax.

A coding agent has no memory between steps, so every step rereads the whole conversation so far, and you pay for it each time. That's why the same small action costs more late in a session than at the start, and why long sessions cost disproportionately more.

For the ideas, read the blog post Spend Less on Agents Without Doing Less: why it happens, a real example, and ten habits that cut the bill without cutting the work. Then use the animations below to see each idea in action, and try the habits yourself in the bill simulator.

Imagine a meeting with one strange rule: before anyone may speak, the minute-taker reads the full minutes aloud, from the very first word. That's roughly how a long agent session is billed. All numbers on this page are illustrative.
  1. 1The minute-taker
  2. 2The triangle
  3. 3The cache
  4. 4Fresh sessions
  5. 5Suitcase of rocks
  6. 6Try it
1

The minute-taker

Before anyone may speak, the full minutes are read aloud, from the very first word.

    Every step rereads the whole conversation; the cost per step climbs through the day.

    2

    It adds up as a triangle

    Each step costs a bit more than the last. The total is the area, not the height.

      Ten steps form a staircase, and the total is the triangle under it.

      3

      The cache discount, and how it breaks

      Caching makes long sessions affordable. It's also easy to lose.

        Without caching, every step pays full price to reread everything.

        4

        One long session vs a fresh session per feature

        Same three features, same work, same quality. One habit apart.

          Feature 1 looks the same in both lanes.

          5

          Suitcase of rocks

          Everything that enters the conversation is carried up every flight of stairs that follows.

            A big build log is carried on every later step.

            6

            Try it: the session bill

            A day of agent work: three unrelated features. Switch habits on and watch the bill.

            Habits
            Illustrative bill for the day $0.00

            An illustrative model, not a price list. The grey track is the default day. Real savings depend on your tools, caching and mix of work.

            Ten habits, by ADLC phase

            The numbers match the ten habits in the post; some help in more than one phase.

            Intent and Plan

            • 2
              Strong model, high effortWhere a good spec saves far more than it costs.
            • 3
              Write it to a fileSpec, rules and decisions a fresh session reads anyway.
            • 4
              Batch your feedbackGroup your comments on a draft into one message.

            Build

            • 1
              One session per featureClear the desk before the next job.
            • 2
              Mid-sized model for routine editsRenames, boilerplate and summaries don't need the strongest model.
            • 7
              Subagents for explorationThey read the thirty files; you get the paragraph.
            • 6
              Lean command outputA PASS/FAIL line, not a 2,000-line log.

            Verify

            • 8
              One command for all checksAnd a check for every recurring mistake.
            • 9
              Keep the independent reviewerCheap next to a long build session, and it catches what the builder missed.

            Review

            • 6
              Short summariesDetails on demand.

            Release and Observe

            • 8
              One post-deploy scriptFor changes that can't affect what users see, the lighter checks.

            Any phase

            • 5
              Don't break the cacheDon't change the rules, tools or model mid-task; finish a thought before a long break.
            • 10
              Watch usage, cap the runawaysLook weekly; limit retries so a confused agent can't loop all afternoon.

            Don't carry yesterday's luggage into today's trip.

            Read the full post: Spend Less on Agents Without Doing Less, with the real example and when a long session is the right call.