Work methods: how the AI solves each job
For every agent you choose how the AI should work: think it through, write and run a script, use a connected system, mix them, or ask you each time.
Updated October 1, 2026 · 3 min read
The five methods
| Method | Good for | Not good for |
|---|---|---|
| Think it through | Judgment, understanding, writing, small or one-off jobs, when you want to read the reasoning | Hundreds of records with the same exact check (slower, more tokens, can slip) |
| Write and run a script | Bulk, repetitive or exact work: many records, calculations, format conversion, fixed-rule checks | Judgment, writing for customers, small one-off jobs, places that can't run code |
| Use a connected system | Fetching or changing data in a system the workflow connects | Work that needs no system |
| Script for the heavy part, thinking for the rest | An exact heavy part followed by interpretation | Simple jobs (it adds steps for nothing) |
| Ask me each time | Jobs that depend on the day: the AI explains the options and waits for your choice | Unattended runs (nobody there to answer) |
You choose. The Co-Pilot suggests one method per Sub-Agent with a one-line reason, tagged Suggested; nothing counts as chosen until you click. Build on the Canvas lists the agents still *Not chosen yet*: keep the suggestions in one click, or review them. Change a method any time in the inspector, under How should the AI solve this?
What a method changes
Only the instructions the agent receives. The inspector shows the exact text for each place the workflow runs, and how many tokens it adds to every call of that agent.
- Think it through adds nothing.
- Write and run a script and Script for the heavy part tell the AI when not to script (a small or one-off job, judgment, nowhere to run code), to keep scripts in
scripts/named after their job and reuse them, to read only a script's summary, and to show a script and wait for your confirmation before running anything that changes, sends or deletes data. - Use a connected system tells it to fetch and change data through the workflow's systems instead of guessing it.
- Ask me each time tells it to describe the ways to do the job with one line of trade-off each, recommend one and wait for your choice.
Where each method works
| Where | What happens |
|---|---|
| Claude Code export | The full instructions. CLAUDE.md gets a short Scripts (work methods) section and the ZIP a scripts/README.md. Claude Code writes and runs the scripts in your environment and asks before it runs commands: the export grants no permissions. |
| Runs on the platform (Live and Simulation) | No code runs here. The script methods work in structured steps (batches, step-by-step totals, results and exceptions only), and Ask me each time picks a way and says why, since nobody can answer during a run. Pre-Flight and the run log say so. |
| OpenAI, Gemini | Filled in from the vendors' docs when their adapters are built. |
AI Context Orchestrator doesn't write or run scripts for your job: it writes the instructions, and the AI does the work in its own environment.
Safety
- A script that changes, sends or deletes data is shown first and runs only after you confirm; when a Human Approval guards the agent, only after that approval.
- Pre-Flight warns when a script agent can reach a connected system and no approval comes before it.
- Credentials stay in environment variables, never in scripts.
Why the Simulation shows no savings
A script can save tokens on bulk work, but only real runs can show it. The Simulation's answers are scripted text, so it shows the flow (*would write scripts/…, run it, read its summary*) and never a savings figure. The inspector shows the one cost known in advance: the tokens the instruction itself adds. Runs of the Claude Code export happen in your environment, so the platform never sees their usage.