Blog · 2026.09.28
Watch the agent think
Most AI coding tools show you the result: a diff, a pull request, a green check. What they don't show you is the hundreds of small decisions it took to get there. Elanous makes those decisions visible.
대부분의 AI 코딩 도구는 결과만 보여 줍니다. 엘라누스는 그 결과에 이르기까지의 판단 — 무엇을, 왜, 어디로 보냈는지 — 를 실시간으로 보여 줍니다.
One run, one constellation
A harness run turns one sentence into a merged pull request: writing the goal, splitting it into pieces, implementing them in isolated workers, running the gates, reviewing the diff unattended, merging. On the Live view each run hangs off the harness as its own branch: the run, its stages, and the pull request it produced. The models it talked to sit on the outer ring.
- The stage the run is in right now pulses, with a ring spreading out from it.
- A stage where something failed turns red.
- Click a run to follow only its decisions and log.
Every decision is tagged with the kind of judgment it was:
When a decision records how many options it weighed, PATHS shows the count.
Decision cards: what, why, where to
Every real decision pops up as a card, then folds away into the place it sent the work. Each card has three parts.
| What it decided | Why | Where it sent the work |
|---|---|---|
| Review found 2 must-fix issues | The reviewer's reason, in one line | → a repair round |
| Gates passed, 12 of 12 | The checks that ran | → review |
| Switched to the next account | That account reached 95%; the budget policy allows it | → the next account |
| Refused a write outside the sandbox | The path it tried to write | → back to the worker, with a way forward |
| Split the goal into 3 pieces | How big the goal was | → 3 workers on the build box |
| Merged #N | Review clean, gates green | → main |
When a card folds, a point of light flies from it to the place it sent the work. Every decision also lands in a stream you can read line by line: [ROUTE] what · why · purpose → where. That is the part we like best: you see not just that the agent acted, but where its judgment went. Missed it? REPLAY plays the last 40 decisions back in order.
When it goes wrong, you see it fix itself
A red pulse on a gate. A card: "gate failed: typecheck in 2 files." An arrow to a repair round. A few seconds later the node turns green and the run carries on. This is the part of Elanous we care about most: it watches itself, repairs what it can, and asks you only when it can't decide.
Two modes, and you choose the cost
- Everyday (default): gauges, stage bars, the decision stream, a model scorecard and review verdicts, built from the summary events Elanous already records. It adds essentially no load.
- Full show (you turn it on, knowing the cost): the decision graph, flying cards and light. To fill the cards, Elanous records more — the reason, purpose and destination of every decision. So it asks once before turning on, lets you pick all runs or just one, switches itself off after 30 minutes, and shows a live cost line (signals per minute, payload size, minutes left) while it is on.
Start one from your phone
Send one line to your Elanous bot on Telegram with /work. The run it starts shows up on the board.
Elanous launches 2026.10.28. See the teaser and how a run goes round on elanous.ai.