Workflows
Workflows your agents follow. Approvals where you decide.
Workflows are the flows in ctxflows, the context layer and the workflows for agent-built software. A workflow is a graph your coding agents follow, agent orchestration you draw once: agent nodes do the work, loops repeat until checks pass, approvals wait for you and triggers start it.
An item reaching the Ready column starts a run, and so does a call to the workflow's webhook. Plan writes the plan, and Build, on the agent you choose, works from it and the rules that apply, repeating until your checks pass. Checks that still fail end the run at Sent back; passing ones wait for you at Review, where Approve ends it at Shipped and Send back at Sent back.
Nodes: agent, approval, condition and loop
Every step of a workflow is one of four kinds of node, so any way your team works can be drawn from the same few parts.
Agent nodes do the work, approvals wait for your decision, conditions pick the path and loops repeat until a result holds. Each node hands the next what was actually produced, never an agent's summary of it.
Loops until checks pass
A loop repeats the nodes inside it until your checks pass, or until the round limit you set is reached.
Checks are your project's own tests, type check and lint, run on your machine by the engine, so an agent's word is never taken for the result. Every round keeps its results, so you see what each attempt changed.
Inputs, outputs and context
Each agent node starts with what it needs, and finishes only once what you asked of it exists.
The context step picks the rules and context that apply before the node starts, so no agent hunts for them or reads them all, and a node can write its result back to a tool you've connected.
Triggers: board column, schedule, webhook and more
A trigger starts a workflow with nobody asking: a board column, a schedule, a webhook, a dataset reaching a condition, a merge to your main branch or a finished milestone.
One workflow holds several, so the same procedure runs from a board, on a schedule or by name with no copies of it, and each trigger sets what its runs may spend.
Your agent picks and runs it
Ask your coding agent for the work in plain words, and it finds the workflow that fits and runs it, node by node.
In your own session it asks you at each approval, and with the engine the same workflow runs unattended, each node on the agent you choose. One set of rules for every agent instead of one per vendor.
Improve from your own runs
A scheduled workflow reads your past runs, what they cost and what slowed or stopped them, and proposes changes to your workflows that you accept or dismiss.
It's a workflow like any other, so you shape it too. Accept a change, and the workflow uses it from its next run; dismiss it, and nothing changes.
From an error report to an approved fix
A new error in production becomes a fix you approved, through a triage workflow and an implementation workflow.
Your error tracker calls the webhook
The new error posts to your triage workflow's address, and a triage run starts with the report.
Triage checks the ledger and files a bug
The run looks the error up in your error ledger, a dataset, so the same error never opens two items. A new one becomes a bug on your triage board, and once it's triaged, it moves to Ready to fix.
The column starts the implementation workflow
The bug reaching Ready to fix starts your implementation workflow. Plan reads the bug, the docs page for the feature it touches and the rules that apply, and writes the plan.
Build loops until the checks pass
On your machine, the engine builds the fix on the bug's own branch and runs your project's checks after every round, until they pass.
An approval asks you
The run waits at its approval with the plan, the check results and the branch, until you decide.
Approve, and the engine lands the change and the bug is done; send it back, and the implementation workflow starts again. Every step stays in the run's history: what each node was given and wrote, which checks passed and who approved.
Questions about workflows
Can a workflow loop until the tests pass?
What starts a workflow?
Can I run a workflow from my own Claude Code session?
Do I need the engine to run workflows?
Can one workflow orchestrate different coding agents, one per node?
Put your agents on a workflow
Draw the way your work gets done once, and change it whenever your work changes.