Give each run a branch

Goal. Run an agent several times without the attempts overwriting each other, then keep the best one. About five minutes.

You need. The CLI and a database with whatever the agent starts from.

Fork per run

strata ./agent branch fork default run-1
strata ./agent --branch run-1 kv put result "answer-1"
strata ./agent --branch run-1 kv put score 7

Each run gets its own branch and writes only there. Ten of them is not a problem: ten forks and twenty writes took 831ms and left the database at 68KB, because forking copies nothing.

strata ./agent branch list

NAME     PARENT   STATUS  GENERATION
default  -        active           1
run-1    default  active           1
run-10   default  active           1

Compare the attempts

Read the same key from each branch:

strata ./agent --branch run-1 --raw kv get score
strata ./agent --branch run-4 --raw kv get score

7
8

Same key, ten answers, none of them interfering. That is the whole point of the pattern: the attempts are isolated by construction rather than by naming convention.

Keep one, discard the rest

strata ./agent branch merge run-4 default

merged run-4 into default: 2 applied, 0 deleted, 0 conflicts (version 55)

strata ./agent branch delete run-1

Deleting the losing branches is optional. They cost almost nothing to keep, and keeping them means you can still answer why the winner won.

The one thing that will bite you

A merge does not carry events or graph data. If the run recorded its tool calls as events on its own branch, merging the winner leaves that log behind, and the receipt still says success:

merged run-1 into default: 1 applied, 0 deleted, 0 conflicts

Two events on the branch. Zero on default afterwards. Nothing warns you.

So decide where the log lives before you start. Record tool activity covers writing it to default while the work happens on a branch, which is usually what you want.

What just happened

A branch is the whole database, so a run branch isolates every model at once: the agent’s key-value state, its JSON documents and its vectors all fork together. There is no per-run namespace to design and no key prefixing to get right.

Failure modes

The merge fails with invalid_argument.engine.branch_point. The branch came from branch create rather than branch fork. Redo the run on a forked branch.

Two runs both merged and the second was refused. They changed the same key. Resolve merge conflicts.

Next

Persist agent memory, for the state that should outlive a single run.