Every agent run on Lattis can be replayed step by step in the console. That promise is cheap to make and expensive to keep: a replay needs the exact inputs, outputs and tool results of every step, stored for 30 days.
Why replays are expensive
A long agent run can pass 40 MB of raw trace. Multiply that by 1.2 million runs a month and the storage bill grows faster than the compute bill. Sampling was not an option: the run you need to debug is always the one you did not keep.
Deduplicate across steps
Most of a step repeats the step before it. We split every step into blocks and store each block once, keyed by its hash.
- System prompts and tool schemas: stored once per run
- Message history: only the new turn is stored per step
- Tool results: deduplicated across runs in the same workspace
That change alone removed 81% of the bytes.
step 14
prompt sha:9f2c (shared)
tools sha:41ab (shared)
history sha:77e0 + 1 new turn
result read_logs, 2.1 KBCompress what is left
The remaining text compresses well with a dictionary trained per workspace, because agents in one workspace call the same tools with similar arguments. Together, the two changes bring a replay to about 3% of the raw trace size.
“The cheapest byte is the one you stored last step.”
What it costs to read
Rebuilding a replay takes 40 ms on average, fast enough to open in the console without a loading state. Retention stays at 30 days by default, and teams that need zero retention can still turn traces off per key.
