Navan grabbed a napkin from the kitchen. An actual paper napkin, the brown recycled kind that tore if you pressed too hard with a ballpoint. He borrowed Jay's pen and started writing numbers.
"Okay," he said. "We have twelve agents running concurrently. Each agent generates roughly two hundred turns per session. Average payload size is about six kilobytes per turn. That's 1.2 megabytes per session, times twelve agents, times—how many sessions per day?"
"Forty," Jay said. "On a busy day."
Navan scribbled. "Forty sessions times 1.2 megabytes is 48 megabytes per day. That's nothing. But now factor in that agents are repetitive. They include the system prompt in every turn. They paste the same error messages. They quote the same documentation blocks."
"How much duplication are we seeing?"
"I ran the numbers last night." Navan flipped the napkin over. "Across a week of agent sessions, 62 percent of blob content is duplicated. Same BLAKE3 hash, same payload. The system prompt alone accounts for 14 percent of total bytes, and it's identical across every single session."
Jay let out a low whistle. "So without CAS, we'd be storing the same system prompt hundreds of times."
"Thousands of times, eventually." Navan underlined a number on the napkin. "With content-addressed storage, each unique blob is stored exactly once. You write a blob, BLAKE3 hashes it, and the store checks whether that hash already exists. If it does, you get back the hash and pay zero storage cost. If it doesn't, the blob gets written once and indexed."
"So the system prompt gets stored once. Every agent that uses it just references the same hash."
"Every agent. Every session. Every turn. One copy." Navan was writing faster now, the napkin filling with arithmetic. "Over a month, at our current usage, we'd store 1.44 gigabytes without CAS. With CAS, based on the 62 percent dedup rate, we store about 550 megabytes. That's a savings of nearly 900 megabytes."
"That's not just storage," Jay said. "That's bandwidth. That's memory. That's every layer of the stack that doesn't have to handle data that already exists."
Navan set the pen down and looked at the napkin. Both sides were covered in numbers, arrows, and one small coffee stain from where his mug had been sitting. "And this is at our current scale. When we're running a hundred agents, the dedup ratio only goes up. More agents means more shared context means more hits in the CAS."
He taped the napkin to the wall next to his monitor. Jay photographed it.
"For posterity," Jay said.
"For the next person who asks why we built a content-addressed store instead of just using a database," Navan replied.
The napkin stayed on that wall for four months. Nobody ever questioned the architecture.
62% deduplication rate is honestly conservative for LLM workloads. I've seen higher. The napkin math checks out.