Useful automation without constant model calls
Updated October 30, 2026. Educational guidance; use current official documentation and your installed version.
The cheapest reliable heartbeat is often ordinary code. Checking whether a message ID, assignment state, or file hash changed usually does not require a language model.
A three-stage budget
| Stage | Work | Model use |
|---|---|---|
| Observe | Read bounded metadata and compare IDs or hashes | None |
| Triage, optional | Classify an ambiguous changed item | Small supported model, short fresh context |
| Produce | Research a question, compose a reply, or review an assessment | Stronger supported model, bounded source packet |
Do not add the triage stage if deterministic rules already identify the next action. A strong model should not reread an unchanged inbox on a timer.
For interactive thinking, execution, composition, and cross-family review across subscriptions, see model routing.
Controls worth implementing
- A small schedule with an explicit timezone and term/project end date.
- Stable source fingerprints and request IDs; deduplicate retries.
- One active batch per workflow; coalesce new evidence rather than starting competitors.
- Daily dispatch and per-agent run caps, plus timeouts and maximum batch sizes.
- Fresh, task-specific context rather than repeatedly carrying an entire chat history.
- No unlimited retry loop, recursive agent cascade, or paid API fallback.
- Change-only notifications; repeated access failures should become one visible blocker.
- Separate count monitoring from assessment judgment. Administrative responses and early submissions are not a past-due grading queue.
Subscription limits are not token price quotes
A subscription-backed agent still consumes the provider's allowance. Model, reasoning effort, context size, tool use, and caching all affect usage. A zero-dollar budget field is not proof of a hard zero-cost limit. Monitor the provider's actual usage dashboard and keep API billing separate.
For current guidance, consult ChatGPT/Codex plan and allowance guidance, separately billed OpenAI API pricing, and API reasoning-effort guidance. Lower reasoning effort can reduce reasoning work, but does not guarantee a particular saving or equal quality.
Review after one week: useful outcomes, failed runs, duplicate work, time saved, and actual allowance consumption. Increase frequency only when those results justify it.
Provider-specific allowances
Microsoft Copilot Chat, GitHub Copilot, Claude Code, and Codex have different allowance and billing rules; do not treat a subscription as unlimited agent capacity. Model routing links current primary guidance and distinguishes Microsoft's conversational chat from GitHub's developer product. Choosing another provider is a deliberate routing decision, not permission to bypass a limit or send data to an unapproved destination.
What a zero in Paperclip means
In Adam's installed October 30, 2026 Paperclip version, subscription-included runs can have token usage and zero recorded marginal spend. The local Codex adapter does not supply a per-run dollar price; the cost ledger keeps subscription-included spend at zero. That is not a claim that the subscription is free, usage is unlimited, or work used no tokens.
Use Paperclip's built-in Costs → Providers quota windows when the provider can report them, plus per-agent token and run counts. Quota windows describe the provider account, not just Paperclip's share, and may be unavailable or delayed. Keep the subscription invoice separate from API cost events. There is no verified simple switch in this installation that turns subscription usage into an API-equivalent dollar estimate; do not relabel guessed estimates as billed spend. A hypothetical estimate would also need the correct model rates and provider-specific cached-token accounting.