In development

Give it a goal. It keeps going.

Your agent shouldn’t
always be waiting
for you.

Give it an outcome and boundaries. Onward is being built to notice what changes, decide what matters, and keep useful work moving. It should return with progress or a specific decision that needs you.

Working today: owner-defined milestones, saved progress, and one complete live goal. Choosing and revising the next step from changing conditions is the next capability.

THE DIRECTION
01 / Your outcomeKeep improving this product.
ObserveWhat changed?
ReasonWhat matters?
Act & checkWhat improved?
02 / What comes backProgress. Or a decision.

An illustration of the intended workflow. The agent revises its next step from what it learns, within the boundaries you set.

A reason to move.
A useful next step.

A heartbeat gives an agent an occasion to look. The missing piece is judgment: deciding what deserves work, choosing the next useful action, and revising course after the result.

01 / SET THE OUTCOME

Tell it what matters.

Define the result you want, the evidence of success, and the limits on what it may change or spend.

02 / KEEP PAYING ATTENTION

Find the useful work.

The intended agent notices changes, investigates their cause, and chooses a next step that moves the goal forward.

03 / RETURN WITH PROGRESS

Make the next decision easier.

Act, check the result, and revise the plan. Bring back verified progress or a specific owner decision when the work needs one.

This is the intended operating loop. Today’s runtime follows owner-defined milestones; automatic planning and replanning are still to be built and validated.

Goals that take days.
Weeks. Months.

The direction is staying with an outcome as evidence changes: keep the context and earlier decisions, revisit the next step, and wait for a useful signal when there is nothing worth doing yet.

ONGOING PAPER RESEARCH

Keep the investigation going.

Carry hypotheses and paper results across weeks. Revisit a test when fresh data arrives, with the owner’s risk limits fixed.

A MONTH OF ONBOARDING IMPROVEMENTS

Build on what each change teaches.

Keep one activation goal across successive fixes. Compare new feedback with earlier evidence and choose the next improvement.

APP HEALTH OVER RELEASES

Follow the repair through.

Keep context from earlier incidents, observe how each release behaves, and return to the goal when new evidence needs attention.

These are intended longer-term workflows. Waiting for data, feedback, or an owner decision should preserve the goal’s context. Sustained operation over weeks or months has not been verified; the published OpenRouter proof covers one bounded goal.

Work that keeps
moving forward.

Illustrative future workflows. These describe the intended product; they are not running Onward services.

PAPER-TRADING RESEARCH

Investigate a drawdown.

A research agent notices a paper-trading drawdown or stale inputs, checks the data, forms a hypothesis, and tests it in simulation. It returns evidence while keeping the owner’s risk limits fixed.

PRODUCT BUILDER

Work through onboarding friction.

When activation drops, an agent reviews where users stop, investigates the likely cause, and prepares and tests a fix. You get the proposed change and the evidence behind it.

APP CARETAKER

Turn failures into repairs.

An agent notices repeated failures, reproduces the problem, and prepares a verified repair. It brings back the result and any specific decision that needs your permission.

A foundation for
that judgment.

Onward builds an optional brief from the goal, current milestone, and recorded evidence. Jev reviews it through three typed decisions: use, rebuild, or stop.

The host assembles the brief. When a valid review is uncertain about that optional summary, the workers receive the full unchanged goal records and facts. An explicit stop or invalid mandatory context still blocks execution.

Jev reviews each candidate against the milestone independently, so both may qualify. The controller still requires its own checks before saving a checkpoint and separate final checks before declaring the whole goal complete.

The current Rust runtime uses two Codex workers with Jev through OpenRouter. Persistent goals run through owner-defined milestones and resume from verified checkpoints. Independent final checks decide completion. Pause, resume, cancel, and export controls are implemented. The service runs in the foreground; no background daemon is installed.

Small core. Concrete evidence.

Built in Rust.
Tested with real workers.

The v3 trial completed one real goal through OpenRouter: two accepted milestones, a separate-process resume, and all 23 final owner checks. Reserved review capacity kept the work within the same 12-call limit.

Read the project evidence ↗
  • One complete live goal

    Both workers passed 15 counting checks. After a separate-process resume, both passed eight CLI checks. The controller accepted the second milestone, saved checkpoint two, and passed all 23 final owner checks.

  • Reviews had room to finish

    All four native worker turns completed. The run used 12 real OpenRouter calls: eight worker requests, two context reviews, and two Jev eligibility reviews. It cost $0.006593 with no unsettled hold and took 24.074 seconds across two controller launches.

  • A small Rust core

    The live-trial binary is about 852 KiB. An earlier version used 1.77 MiB idle and peaked at 3.88 MiB active in one Mac fixture with fake providers. Codex workers use additional memory: 147.73 MiB combined within a 203.80 MiB sampled process tree. Those historical measurements do not describe the v3 live run.

  • Validation still ahead

    One bounded goal establishes live completion and resume. Automatic planning and replanning are not implemented; broader coding capability and 24-hour reliability remain unverified. Earlier native fixtures cover crash recovery and rejection of failed checks.