Zero Token-Waste: How akiflow Keeps Token Cost Low Without RAG or Daemons
akiflow runs its agent council on direct file reads and grep-based log slicing: no vector DB, no background process. Why every read into context is billed again on every later turn.
Where does multi-agent orchestration spend tokens, and how do you avoid it? Many multi-agent frameworks add infrastructure: background processes, a vector DB, embedding pipelines. akiflow goes the other way: everything is a file on disk plus a few Python scripts that only read or write files, run when you invoke the skill, and stop when it finishes.
This article describes where tokens actually burn in a long session and the mechanisms akiflow uses to keep cost down.
Where do the tokens go?
Every turn re-sends the whole history. Anything already in context is billed again on every later turn: a read of size S at turn t of a T-turn run costs roughly S × (T − t). Pulling a 50k-token room into context at turn 50 of 200 costs about 7.5 million cache-read tokens from a single read call.
Measured on a real run of akiflow itself (recorded in the repo's docs/arch/akiflow.md): even with three aki-maker workers doing every file edit, the lead still held 70% of cache-read and 66% of output. Moving the work elsewhere did not cut the cost; the expensive part is the lead's own accumulated context.
The mechanisms that keep cost low
- ◆Direct file reads: rules are markdown files in
~/.aki/akidevrule/, and the lead points each agent at exactly the files it needs bytopic.A1address. There is no embedding step and no semantic search. - ◆Room-log slicing:
council_read.pylocates with--grep(matching lines only, each tagged with its turn), then pulls exactly the needed turn with--turn. It also offers--index,--stats(turns and bytes per agent), and--tail. - ◆Mechanical work goes to a cheap tier:
aki-handshas only Read, Grep, and Glob, may not judge, and runs on the cheapest tier that can do the job. The lead does not do menial work itself. - ◆No background process: each run has a workspace under
~/.aki/agent-council/<project>/, pruned after 30 days. Outside that folder akiflow leaves no state. - ◆One-batch convening: agents are spawned together and share the same rule-file prefix, so the prompt cache stays warm. An agent spawned late pays a colder read.
- ◆On only with a reason: under law R2, every seat and check is off by default and turns on only when the run produces a reason.
- ◆Close-out accounting:
council_cost.pytotals the exact Claude-side tokens of the lead and every subagent, so actual cost is reconciled against the roster declared before the run.
Limits worth knowing
This describes mechanisms, not a savings figure: the repo publishes no percentage reduction against other frameworks. council_cost.py measures only the Claude side. A lane run through another CLI (for example agy headless) draws on a different vendor's quota and is reported beside the table, not added into it.
A council still costs more than working directly. Work that does not pass the activation gate should just be done directly; do not open a room.