I handed Photon my brief — can I steer the build while it's running, or do I just wait for delivery?
4 min read
The short answer: you cannot push new instructions into a running build, but the window is not a black box either. Photon's own page — updated November 16, 2026 — describes a defined middle: you watch the plan and each agent's status in the workspace, a Watcher automatically steers any task that stalls, repeats itself or drifts from the request, and Photon asks you directly when a decision is yours to make, with a browser notification when a question is waiting or the work is finished. What the page does not describe is a way to inject new instructions mid-build, pause an agent, or edit the plan after it is set. If you need to change direction, that change lands as a new request in the workspace — which is how post-launch changes already work — not as a live edit to the running build.
Source: Photon AI Coding Agent — Describe It, Get It Built
What the page establishes about the in-build window
The workflow runs like this: you describe the result you want, a lead agent plans the work and assembles a team of AI agents, and the agents write, run and check the work in your workspace while you watch the plan and each agent's status. The Watcher is the part that does steering for you — it checks every running task and corrects the course when one stalls, loops, or drifts from what you asked for. You are not expected to notice those failures yourself.
Source: Photon AI Coding Agent — Describe It, Get It Built
Your part during the build is mostly reactive: answer when Photon asks, because the notification exists for exactly that moment — a decision that belongs to you, waiting. Every account has its own workspace and its files stay there between sessions, so the build is never running somewhere you cannot look. Finished work arrives in Results, where you open it, preview it, download it or publish it, and a page stays visible only to you until you publish.
Source: Photon AI Coding Agent — Describe It, Get It Built
The steering surface is front-loaded
Since live intervention is not part of the described workflow, the levers that do exist sit before the build starts. The brief itself is the first one. The page also lists team instructions and reusable skills that the lead agent and every worker follow, plus a daily, weekly or monthly spending limit. Those are set up ahead of time, and they are where your preferences actually bind the agents. One sentence is enough to start, and you add files or details when you have them — but the more checkable the brief, the less you will wish you could have interrupted.
Source: Photon AI Coding Agent — Describe It, Get It Built
A proposed technique: name your acceptance milestones in the brief
This next part is my suggestion, not something the page describes. Write two or three named acceptance milestones into the brief so the build has defined moments you can inspect and redirect against before hand-off. Instead of "build me a booking site," the brief names checkpoints — "working booking form with validation," "calendar view showing real times," "final pass on phone widths" — each one written so you can test it end to end.
The milestones do two things. First, they give you named checkpoints to hold the delivered work against: you verify each one individually instead of making one all-or-nothing judgment about the whole project. Second, because the Watcher steers tasks that drift from the request, the milestones give that drift check concrete reference points rather than a single vague ask. The page does not claim intermediate results appear in Results before delivery, so treat the inspection moments as arriving with the work, not as live windows into it. A milestone that says "the form rejects a past date" is checkable in the private preview; one that says "it should feel professional" is not.
Source: Photon AI Coding Agent — Describe It, Get It Built
Checking the claims in this article
Everything above is verbatim from Photon's product page: the watchable plan and agent statuses, the Watcher's stall-and-drift steering, the decision prompts with browser notification, the Results hand-off, and the absence of any described mid-build injection, pause or plan-editing control. That page is SentX's first-party disclosure — the company describing its own product — not independent certification. The model card's known-limits habit applies directly here: open the cited source and check the supporting passage. The page is one you can open, and every workflow claim in this article sits in its text.
Source: Photon AI Coding Agent — Describe It, Get It Built; Victoria AI Model Card | Capabilities, Evidence and Limits
So the practical version: write the milestones before you hand over the brief, keep an eye on statuses, answer when Photon asks, and check against your named milestones when the work lands in Results. That is the whole steering surface, and it is enough.
Source: Photon AI Coding Agent — Describe It, Get It Built
Sources
- Photon AI Coding Agent — Describe It, Get It Built — SentX SI
Supported claims and limits
- Page updated November 16, 2026
- Lead agent plans the work and assembles a team of AI agents
- Agents write, run and check the work in the user's workspace while the user watches the plan and each agent's status
- A Watcher checks every running task and steers it when it stalls, repeats itself or drifts from the request
- Finished work arrives in Results, where the user opens, previews, downloads or publishes it
- Photon asks when a decision is the user's to make, and can send a browser notification when a question is waiting or the work is finished
- Every account has its own workspace, and its files stay there between sessions
- A page is visible only to the user until they publish it
- Write team instructions and reusable skills that the lead and every worker follow
- Set a daily, weekly or monthly spending limit
- One sentence is enough to start; add files or details when you have them
- The page does not describe a mechanism for injecting new instructions mid-build, pausing agents, or editing the plan after it is set
- Victoria AI Model Card | Capabilities, Evidence and Limits — SentX SI
Supported claims and limits
- The card is a first-party product disclosure, not an independent certification
- Known limits: answers can contain incorrect facts, citations or calculations; open cited sources and check the supporting passage