Durable State

Durable state is the workbench that survives the current turn. Plan files, checkpoints, task state, session summaries, logs of what's been done, diffs, memory stores — everything that lets work exist outside the model's fleeting attention.

Context management vs. durable state

These two primitives are easy to conflate, and the distinction is the whole point:

Context management        decides what gets brought INTO the model right now
Durable state             preserves facts, artifacts and progress OUTSIDE
                          the prompt, so they survive the current run

One is a spotlight; the other is the warehouse the spotlight sweeps across.

What durable state looks like

For a coding agent: a branch with changes, a test log, a plan document, a summary of what failed on the last attempt.

.agent/
├── plan.md            # steps, with 3 of 7 checked off
├── attempts.log       # attempt 2: migration failed — FK constraint
├── checkpoint.json    # resume from step 4, branch: fix/session-bug
└── notes.md           # "the flaky test is unrelated — ignore it"

For a research agent: a source map, extracted quotes, a confidence table.

In the wild

Plan files that agents write before touching code. Task checkpoints that survive a crash. The session summary a harness saves when a conversation is compacted. Git branches as recoverable state. Memory files that carry a lesson from one session into the next. All of it is this primitive.

State preserves — it does not coordinate

A plan file can say what should happen; it cannot decide what to retry, or when. A log can record that a command failed; it cannot decide whether that failure warrants asking a human. A memory can preserve a lesson; it cannot schedule the next run.

Once a workflow has a lifecycle — start here, pause there, resume after this, ask before the risky step — state alone is clearly not enough. Something has to move the work.