We asked AI to take work off our plates. It did. Then it put a new kind of work back on them.
Open enough agent threads and the pattern becomes obvious. Each one can research, write, analyze, or build faster than you could alone. But every thread still needs context. Every output needs checking. Every failure needs diagnosis. Every exception needs a decision.
Soon, your day becomes a loop: start a thread, explain the job, check its progress, correct the course, recover lost context, review the output, and start again.
The agent is faster. You are busier.
That cannot be the future of knowledge work.
More execution is not the same as more leverage
Most AI tools make it easy to start work. That matters. But starting more work is not the same as scaling yourself.
If your attention has to grow with the number of threads, the extra output has not created much leverage. It has turned the human into a dispatcher, status checker, and full-time supervisor.
The problem gets worse when the work repeats.
A one-off conversation can live in chat. You explore, adjust, and move on. But a process you run every week should not depend on remembering the right prompt, rebuilding the same context, and personally watching every step.
Repeated work needs a durable method.
What are the stages? What does each stage need? What counts as good evidence? Where should an exception go? When should the work stop? Which decisions still belong to a person?
From chat threads to agent systems
AI adoption starts in chat. The next meaningful move is not opening more chats. It is changing the unit of work from a conversation into a system.
Slipform is a Mac app for turning repeatable agent work into explicit workflows. Codex is the supported runtime today.
You define the stages, required evidence, evaluation criteria, and routes the work can take.
A research workflow might gather sources, build a thesis, challenge its claims, review its risks, and return a decision packet. You can author a route that sends weak evidence back to research, moves unresolved risks to review, and brings the final decision to you.
The workflow holds the method. Each run creates a record.
That record includes the artifacts, evaluations, performance, token use, route taken, and final decision. You can inspect what happened without reconstructing it from a long conversation. As runs accumulate, you can compare them, find the weak stage, and decide what to revise.
The questions change.
- Where is the workflow spending time and tokens?
- Which evaluation is consistently weak?
- What caused the rework?
- Did the exception follow the right route?
- What should change before the next run?
That is a calmer relationship with AI. You are not removed from the work. You are working where your attention matters most.
The human owns improvement
Slipform does not decide what good means.
It does not approve work for you or rewrite its own workflow. At a human review gate, you choose whether to approve the run, request rework, or fail it.
The record helps you make that decision. It can also show you where the workflow needs attention. But revision remains yours.
A workflow that changes itself without your judgment may become different without becoming better. Improvement needs an owner.
The goal is not to supervise fewer threads by trusting them blindly. It is to replace scattered supervision with a method you can inspect.
Codex is the starting point
Codex is the only supported runtime in the current beta.
That is the starting point, not the full direction. The durable object should be the workflow: your stages, standards, routes, evaluations, and run history. The execution layer will continue to change.
We are designing Slipform to extend beyond a single model or provider. Broader runtime support is not shipped today.
The current product is concrete: repeatable Codex work, authored as workflows, recorded as runs, and governed by human decisions.