About
We kept explaining the same thing to the same agent.
Agent Flow grew out of directing AI on real projects.
Powerful coding agents need clear rules, visible decisions, and a memory of what went wrong. None of those live inside the model.
Directing an agent is not the same job as writing code, and it was landing in the gaps between tools. Scope drifted quietly. A rule that mattered lived in somebody's head. The lesson from Tuesday was gone by Thursday, because the session that learned it had ended.
The team behind Agent Flow
Why this
exists.
That experience shapes how we think about AI development tools. The code matters. So do the instructions, the decisions and the reviews that make it possible to trust the work.
Where we put the control.
Three ideas sit at the centre of the product, and we are explicit about where the two providers behave differently.
Check the action before it happens.
For Claude Code, governed tool calls are checked before they execute. A rule can deny a tool, a restricted path, or a file change beyond the session budget while the action is still a proposal.
Codex uses a configured sandbox for preventive limits, so Agent Flow checks its activity as it is reported, after an action has started. Six session modes set the intended reach, with enforcement that depends on the provider.
Govern both providers the same way.
Run Claude Code and Codex side by side in one workspace. Shared governance rules, a session ledger and plan review give you one way to direct the work, whichever provider is doing it.
Model and reasoning effort are chosen separately for coding, research, planning, debugging and review, per provider.
Carry the lesson into the next session.
With learning enabled, detected failures and governance violations become saved project lessons. Relevant ones are selected for later sessions and handovers, within the available context budget.
A repository codebase map and planning documents hold the project's context too, so the next task starts from what is already known.
What we are careful not to claim.
- The two providers enforce differently, and we say so on the page rather than in a footnote.
- Review marks record that you read something. They do not revert a change or certify it.
- A connector appearing in the workspace is not proof of a live connection.
- Your AI provider costs stay yours. Agent Flow is the workspace, not the model.
Your judgement. In the workflow.
You decide the scope, the team and the rules. Agent Flow brings those decisions into the session and keeps the work where you can see it.