Quick Summary
- Automate from current source files and runtime evidence.
- Let GPT-5.6 Sol through Codex carry bounded execution volume.
- Let Claude Fable 5 carry judgment and voice.
- Gate external side effects at the capability layer.
The Automation Contract
Claude Code and Codex CLI can read files, call configured tools, run commands, and update local artifacts. That capability becomes useful when the task starts with a written contract.
The contract names:
- allowed inputs and output paths
- deterministic eligibility and suppression rules
- the model judgment required
- the external effects that remain gated
- success, no-work, and failure outputs
Without those boundaries, an agent is a flexible demo. With them, it can become a repeatable worker.
File and Data Workflows
File work is the safest place to start because the result is inspectable.
Good first automations include:
- classify and route incoming files
- normalize CSV or JSON records
- search a knowledge base and assemble a source packet
- update a report from current logs
- rebuild generated HTML from Markdown sources
- compare runtime output with documented claims
GPT-5.6 Sol through Codex carries this execution volume well when the acceptance criteria are explicit.
Browser and API Workflows
Browser and API access expand the blast radius.
Read-only inspection can gather evidence. Writes need stronger controls:
- resolve the exact account and target
- show the outbound text or mutation
- verify identity, authentication, and capability
- use dry-run and idempotency controls where available
- retain the response and failure receipt
The prompt does not substitute for a capability gate.
Content Workflows
A source-grounded content workflow separates research, drafting, judgment, and publication.
One useful split is:
- Sol searches and extracts from confirmed sources.
- Sol structures the brief and checks required fields.
- Fable judges the angle, voice, and persuasive choices.
- Sol formats, builds, and runs deterministic checks.
- A person approves consequential claims and publication.
The memory plane keeps the source packet, accepted copy, corrections, and delivery receipt.
Human Gates
Keep a fresh human gate for:
- sends and publishing
- purchases, charges, refunds, or payouts
- destructive actions
- activation from off to on
- unresolved recipients, accounts, or material claims
- scope changes that create a new external effect
Bounded local file changes do not need the same ceremony because the diff is visible and reversible.
Verification
Test new automation foreground on a tiny batch.
Read the actual output. Check the diff. Confirm that no protected files or unrelated user changes moved. Then scale.
Every automated run should leave an observable result:
- artifact written
- explicit no-work result
- receipt with identifiers and counts
- failure naming what it inspected
The rule is operational: runtime truth beats documentation, and code present is not code deployed.
Related: