Working With an AI Coding Agent as a Developer: How to Stay in Control
4 min read
A developer's worry about an AI coding agent is rarely whether it can write code. It is what happens when it writes a lot of code, quickly, while you are busy with something else. The answer is the one you would give for any capable new team: agree the rules up front, give the work a defined place to happen, read what was produced, and cap what it may spend. Photon, the AI coding agent from SentX SI, has a control for each of those, and this is how they fit together.
It is a team, so manage it like one
Photon does not work as one assistant answering in one thread. You describe the result you want, a lead agent plans the work and assembles a team of AI agents for it, and the agents write, run and check the work in your workspace. You can watch the plan and each agent's status while that happens.
Source: Photon: the AI coding agent
That structure tells you where your effort belongs. Nobody manages a team by dictating every line. You agree how the team works, and then you inspect the output. The same division applies here.
Write the rules once
Team instructions are the first control worth setting. You write them once, and the lead and every worker follow them. They are the place for everything you would otherwise repeat in each request: the language and tone of reports, what is in scope and what is not, how finished work gets checked before it is handed over.
Reusable skills sit beside them. A skill is a set of instructions the team can call on by name, which suits procedures you want done the same way every time. A team that ships small internal tools might keep one skill for how a tool is tested before delivery and another for how its README is laid out. Those two are examples, not built-in skills.
Good standing rules are short and checkable. "Run the tests and report the result before marking a task finished" can be verified. "Write clean code" cannot.
Give the work a real place to run
By default the work happens in your workspace. Every account has its own, and its files stay there between sessions, so the second task starts from what the first one left behind.
When the job belongs somewhere else, you can connect your own server so Photon can work on it, and open a terminal to it from the workspace. That changes the kind of task you can hand over, from producing files to working where your software actually lives. It also raises the stakes, so the ordinary discipline of a shared environment applies: begin with a server you can afford to break, keep your own access separate, and widen the scope as the results earn it.
Source: Photon: the AI coding agent
Read what it wrote
Nothing about an agent removes code review. You can read every file the agents wrote, and the layout makes that practical: finished work arrives in Results, where you open it, preview it, download it or publish it, apart from the drafts and scripts produced on the way.
Supervision does not rest on you alone. A Watcher checks every running task and steers it when it stalls, repeats itself or drifts from the request. That covers the familiar failure of long agent runs, the loop that makes no progress. It is a safety net for the process and no replacement for your judgment of the result.
Photon also stops to ask. It asks when a decision is yours to make, and can send a browser notification when a question is waiting or the work is finished, so a long task does not need you sitting in front of it.
Cap the spend before you start
A team of agents working through a real task uses more than a single chat reply does. Two things keep that predictable. You can set a daily, weekly or monthly spending limit. And plan usage resets every week: if you reach a limit, you keep working from your wallet balance or wait for the reset.
Set the limit before a long task, not after a surprising one.
Choose the plan by the work
The plans differ in ways a developer will notice. Plus runs the standard Photon model. Pro adds the Pro model and ten times the usage of Plus. Ultra gives twenty or fifty times the usage of Plus and is the only plan with Max thinking mode. Published sites get an automatic name on sentx.site on Plus and Pro, and a name you choose on Ultra. Plus costs USD 20 a month, Pro USD 100, and Ultra USD 200 or USD 500.
Source: Photon: the AI coding agent
A sensible way to decide is to run one real task on the one-time trial that a new account without a paid plan gets, see how much of your daily work you would hand over, and size the plan to that.
Sources
- Photon: the AI coding agent — SentX SI
Supported claims and limits
- A lead agent plans the work and assembles a team of AI agents that write, run and check the work in the account's workspace.
- Team instructions and reusable skills are followed by the lead and every worker.
- Every account has its own workspace whose files stay between sessions; a user can connect their own server and open a terminal to it.
- Every file the agents wrote can be read, and finished work arrives in Results.
- A Watcher checks every running task and steers it when it stalls, repeats itself or drifts from the request.
- A daily, weekly or monthly spending limit can be set; plan usage resets every week and work can continue from the wallet balance.
- Plus is USD 20 a month with the standard model, Pro is USD 100 with the Pro model and 10x the usage of Plus, and Ultra is USD 200 or USD 500 with 20x or 50x the usage of Plus and Max thinking mode.