You have a goal. Your agent finishes a task. Then you’re back to figuring out what it should do next.
Onward is being built to carry that work forward over days, weeks, and months: assess what changed, choose a useful next step, and follow through while you’re away.
Today: it executes milestones you define. Next: choosing and revising the plan itself.
PROGRESS IS THE POINT01 — ∞
DESIGNED TO KEEP MOVINGBRING BACK PROGRESS
A goal that stays in viewUseful next stepsEvidence of progressRoom to change course
01 / In practice
Less directing. More moving forward.
The goal is to give your agent an outcome and boundaries once, then come back to useful work and the reasoning behind it.
Illustrative workflows · the planning capability is still to be built
A trading bot · paper research
“Build evidence for this trading idea.”
Over a series of paper experiments, the agent could check the data, test assumptions, track what holds up, and choose the next research question as conditions change.
You come back to
A growing research record: what has been tested, what changed, and what the evidence supports so far.
Research with paper results, within fixed limits. Live trading is outside this example.
A product you’re building
“Keep improving onboarding this month.”
The agent could find where people get stuck, make an authorized improvement, wait for useful feedback, and use what it learns to choose the next change.
You come back to
A sequence of checked improvements, the evidence behind each, and a plan that develops with the product.
Changes and experiments stay within the access and release permissions you set.
An app that needs care
“Keep this app healthy as it grows.”
Across releases, the agent could investigate recurring failures, verify repairs, retain what it learned, and notice when a previous solution stops working.
You come back to
A history of verified repairs, unresolved risks, and the next issue worth your attention.
It asks when the next action needs authority you haven’t given it.
02 / Longer-term goals
Goals that take days. Weeks. Months.
Some outcomes need many attempts, time for results, and decisions you can’t write in advance. Onward’s ambition is to keep the goal and what it learns connected through all of them.
One goal: improve onboarding over the next month
Illustrative future workflow · not a completed run or a fixed schedule
At the start
Understand the outcome.
Establish what better means, inspect the current experience, and choose a first step within your limits.
As work progresses
Carry the learning forward.
Check a small improvement and retain the result, the reasoning, and the questions still open.
When evidence arrives
Revise the approach.
Use real feedback to keep a useful change, investigate a weak result, or pick a different next experiment.
When you return
Know where things stand.
See what improved, what remains uncertain, and the next decision. The goal stays in view between conversations.
Progress can include waiting for a result. The intended agent keeps a clear reason to wait and resumes when useful evidence arrives, within the permissions and budget you set.
This is the longer-term product direction. Today’s prototype executes owner-defined milestones; choosing and revising its own plan and reliable operation over days or weeks remain to be built and demonstrated.
03 / Taking initiative
A check-in should lead to a decision.
A heartbeat creates a moment to look around. The useful part is what the agent thinks through: what changed, what matters now, and what it should do next.
The loop we’re building towardInteractive illustration · no agents running
Keep the whole goal in view.
What changed? What matters now?
Read the goal, the progress so far, and the latest evidence. Carry forward the context needed to make a useful decision—even after a pause or a fresh agent turn.
Intended behavior: understand the situation before choosing work
Taking initiative also means knowing when to wait. If there’s no useful action, it should say what it’s waiting for. If it needs you, it should bring a specific decision.
04 / Current build
The foundation works. Initiative comes next.
Today’s prototype follows milestones you define. One small real coding goal saved its progress, resumed in a fresh process, and completed.
Working today
Execute and check owner-defined milestones
Save verified progress and resume it
Keep the goal, boundaries, and facts in context
Enforce permissions and a shared budget
The next proof
Start with an outcome and no supplied task list
Choose bounded steps toward that outcome
Revise the plan when new evidence changes it
Verify progress without another owner prompt
Automatic planning and replanning are not implemented yet. Continuous 24-hour reliability remains unverified. The next test needs to show useful judgment under the same permissions and budget.
Inside the current build Real trial, architecture & limits
Two isolated Codex workers propose work through OpenRouter. Jev reviews context and assesses candidates. The Rust host enforces limits and independently checks the result. This is the execution foundation for the planned decision loop.
LIVE GOAL COMPLETEDPERSISTENT GOAL / OCT 2026
24.1s
Across both process launches
$0.006593
Total provider cost · zero holds
23/23
Independent final checks passed
⊣
Verified, saved, resumed, and completed. Both workers passed 15 counting checks, then eight CLI checks after a separate-process resume. Jev reviewed both milestones, and the host ran all 23 final checks. The fixed 12-request allowance now protects mediation and gives each worker its own share.
GOAL COMPLETE
All four worker turns used live models. Both workers implemented the CLI early; the selected second milestone retained that code while running its own checks and review. This is one small completed goal, not proof of broad capability or 24-hour reliability.
↗
The goal survives an uncertain summary. The host builds a brief from verified progress. Jev reviews it; an uncertain non-stop answer uses the full validated facts. An explicit stop still blocks. Each candidate is assessed independently, and host verification remains required. Jev makes structured decisions; it does not write prose summaries.
Research preview. Request capacity is protected before dispatch: each worker gets a bounded allowance, and Jev keeps its review slot. The live trial used exactly 12 requests with zero unsettled charges; separate engineering tests cover local quota exhaustion and crash handling. Earlier outcomes remain recorded: the original pipeline stopped at its preference gate, the first context policy stopped before workers, and v2 saved one checkpoint but exhausted its requests before the second review. The cancelled fixture soak proves no 24-hour capability. Automatic planning, a permanent background service, and continuous real-provider operation remain ahead. Onward is an independent project.
05 / Small by design
Built to stay out of the way.
A tiny Rust controller is the foundation for keeping work available with little overhead. The agents and their tools use more memory; the controller is just one part.
852KiB
Live-trial release binary
3.9MiB
Prior fixture controller peak
The live-trial binary is 872,720 bytes. Earlier fixture measurements used a prior runtime: 1.8 MiB idle, ~148 MiB for both Codex workers, and ~204 MiB for the full process tree. Live memory and 24-hour reliability remain unmeasured.
The direction is onward.
Work should continue after the conversation.
Give it the outcome and the boundaries. Come back to progress and the decisions that got it there.