Ralph loops

A Ralph loop runs one objective as a chain of fresh rounds: each round starts a new agent with no memory of your chat, works, and reports whether to carry on. DataMind Oda starts a loop only when you ask for one, and the loop holds the turn until a round reports the work done or the rounds run out.

Ask for a Ralph loop

1. Say what to complete, and ask for a Ralph loop, or a fresh-agent iteration, in your own words. Oda reaches for a loop only on that explicit ask; it never starts one by itself.
2. Send the message. The loop appears as a card titled ralph that shows your objective, and the turn it runs in does not finish until the whole loop settles.

• The objective is immutable: every round receives the same one. To work toward something else, stop the loop and ask again.
• The model's own guidance reads: Use the ralph tool ONLY when the direct human explicitly asks for a Ralph loop or fresh-agent iterative execution.
• A loop suits work that benefits from starting each attempt clean. For ordinary long-running objectives Oda uses Goals, and the Relentless preset keeps working until the task is done or it needs a decision; for bounded delegation and fan-out it uses Subagents or Workflows.
• The loop sits behind the Ralph autonomous loop feature: turn it on in Settings > Developer > Advanced features (Advanced features catalog). In a Relentless chat it works without the switch, because choosing the preset is the opt-in; an organisation policy that holds the switch off, or a feature the computer cannot run, still keeps it off.

What a loop call carries

The call supplies data only. The loop itself, the route to the child agents, the shape of the round reports and the handoff checks are fixed by the deployment, and a call cannot change them.

What happens in each round

• One fresh agent. A round starts a child agent with no conversation seed and no history of the chat.
• A bounded handoff. Only the last round's closing report crosses to the next round, capped at 16,384 characters. The shared workspace is the only other long-term memory the rounds have.
• The chat's folder. Every round works in the folder the chat works in.
• A closing report. Each round ends with a structured report: continue (next steps, no blocker), complete (evidence, no next steps, no blocker) or blocked (a concrete blocker). A report that is missing, does not fit its state, or runs over the handoff cap ends the run.

How a loop ends

• The loop ends when a round reports the work complete, when a round reports a blocker, or when the round limit is reached. It settles as complete, blocked or budget-limited, and the result carries the last bounded report and how many rounds ran.
• When a round reports the work complete: Ralph worker reported completion after <n>.
• When a round reports a blocker: Ralph worker reported a blocker after <n>.
• When the round limit is reached: Ralph reached its <n> limit; the worker reported work remaining.
• In these lines <n> reads as the number of rounds, such as 1 round or 2 rounds.
• A round that ends without its report stops the run as well: the result reads Ralph round <n> child failed before producing a structured report. and carries Last successful handoff: with the last report, or No previous handoff was available. when there is none. A round failure is terminal for the loop.
• Stopping the turn that runs the loop cancels it, and the result reads Ralph workflow was cancelled. A loop that fails gives Ralph workflow failed: followed by the failure.

Follow a running loop

• The chat shows the loop's card and, when the loop settles, one closing result. The rounds' inner reports do not appear in the chat.
• In Activity, the loop folds into a single Ralph row with each round as a member (Activity). Its records are observational: they describe the run, and they never reach a model request.

The limits behind a loop

• 256 rounds is the default and the ceiling for one run: a call can ask for fewer rounds with maxRounds, but not more. The round count is the only bound on a loop's total effort; no elapsed-time deadline stops it.
• Reports and the final result are each capped at 16,384 characters by default. A longer final result is cut with a … [truncated] notice.
• One round starts exactly one fresh agent: a loop does not fan out inside a round.
• A loop is foreground only: no background job, no job id and no resume. For work that can run in the background, see Background jobs.
• Completion and blockers are the rounds' own reports, not an independent check of the work: Completion and blockers are worker reports, not independent evaluation.
• An ordinary child failure is terminal. Where a workflow script carries on past a failed helper, a Ralph run does not.
• Which provider serves the fresh rounds is a deployment setting; a call cannot override it.

Why a loop is refused

The call is checked before any round starts, and a loop that fails a check starts nothing. Each refusal names its reason:

• the objective is empty: Ralph objective must be a non-empty string;
• the round cap is not a positive whole number: Ralph maxRounds must be a positive safe integer;
• the round cap passes the ceiling: Ralph maxRounds <value> exceeds the deployment ceiling <ceiling>;
• there is no chat behind the call: Ralph tool requires a calling agent (exec.agent was undefined);
• the child-agent route is not registered: Ralph subagent provider "<name>" is not registered;
• it cannot produce structured reports: Ralph subagent provider "<name>" does not support structured output;
• it would inherit the chat's context instead of starting fresh: Ralph subagent provider "<name>" inherits parent context; Ralph requires a fresh provider.

Questions, answered

Does Oda ever start a Ralph loop on its own?

No. Oda starts a loop only when you explicitly ask for one. Everything else stays with its usual tools: goals for long-running objectives, subagents and workflows for delegation and fan-out.

What is the ralph card in my chat?

It is the loop's call, shown as a card. It carries the objective the loop works toward and holds the turn until the whole loop settles.

How do the rounds pass work between them?

Through a bounded handoff and the workspace. Every round starts fresh, with no conversation seed and no history of the chat; only the last round's closing report crosses to the next round, capped at 16,384 characters. The shared workspace is the only other long-term memory the rounds share.

Do I see what each round did?

Not in the chat. The chat shows the loop's card and one closing result; the rounds' inner reports do not appear there. In Activity, the loop is a single Ralph row with each round as a member.

How long can a loop keep going?

The round count bounds a loop, not the clock: no elapsed-time deadline stops it, and 256 rounds is the default and the ceiling for one run. A call can set a smaller maxRounds, but a cap over the ceiling is refused: Ralph maxRounds <value> exceeds the deployment ceiling <ceiling>.

What happens when a round reports a blocker?

The loop stops. The result reads Ralph worker reported a blocker after <n>., and it carries the last report. The loop does not continue on its own.

What does the budget-limited ending mean?

The loop reached its round limit while the last report still had work remaining. The result reads Ralph reached its <n> limit; the worker reported work remaining., and it carries the last bounded report and the rounds that ran.

Can I change the objective while the loop runs?

No. The objective is immutable: every round receives the same one, and there is no way to edit it while the loop runs. Stop the loop and ask again with the new objective.

Can I stop a running Ralph loop?

Yes. The loop holds the turn it runs in, so stopping that turn cancels it (Stopping Oda and queued messages), and the result reads Ralph workflow was cancelled.

Can a Ralph loop run in the background?

No. A loop holds the turn until the whole loop settles, and it has no job id, no background start and no resume. The workflow tool is the tool whose runs can start as background jobs.

Why can't Oda run the loop I asked for?

A check failed before any round started: an empty objective, a round cap that is not a positive whole number or passes the ceiling, or a child-agent route that is not registered, cannot produce structured reports, or would start from the chat's own context. A refused loop starts nothing, and the refusal names its reason (Why a loop is refused).