Bring the account that owns the work
Forge does not need to resell every model token. Where a provider supports personal-subscription authentication, Forge can run through that account; provider API credentials and organization-managed access remain separate lanes. Today, personal ChatGPT subscription authentication powers Codex turns. Additional providers are enabled only after their authentication and isolation contracts are certified.
That distinction matters. A personal subscription, an API key, and an organization allowance have different owners, limits, and security boundaries. Forge should always show which lane will pay for a turn before work begins, and it should never silently spend a personal account when organization funding was expected.
Route by work, not by brand loyalty
The most expensive model is not automatically the best model for every step. Use a capable lead for high-judgment planning, a fast model for mechanical discovery, and deterministic tools for checks that do not need a model at all.
A good routing policy preserves quality where judgment matters and removes model calls where a compiler, test runner, linter, or source lookup can answer with more certainty. The aim is not the lowest token count; it is the lowest cost of a trustworthy outcome.
- Use one lead to hold the full objective and decision history.
- Delegate only bounded, independent tasks.
- Run deterministic checks before model-based review.
- Avoid repeated reviews over the same unchanged diff.
Make budgets part of the work packet
Forge task packets can carry an autonomy budget such as an implementation-attempt cap. Repository policy can also bound active agents, model tiers, and the total token budget for a task tree. Those limits are operational controls, not vague reminders buried in a prompt.
When a limit is reached, the correct behavior is to compact the evidence and ask for a decision. Do not quietly downgrade required verification or keep generating increasingly speculative attempts. Safety, correctness, and proof gates remain intact even when the execution strategy changes.
- 01
Set the boundary
Choose the active-agent cap, model tier, attempt cap, and ship authority.
- 02
Measure useful work
Track completed checks, evidence, and changed outcomes, not raw activity.
- 03
Escalate deliberately
Ask before expanding scope, authority, or the approved usage envelope.
Commands make the policy repeatable
Forge's command and skill loops turn policy into a repeatable workflow. A task packet establishes the scope and autonomy budget; focused implementation and QA loops define their gates; shipping remains a separate boundary. Teams can encode stricter defaults in repository instructions so every agent starts with the same cost and quality expectations.
A complete organization-level budget ledger and policy surface is still being developed. Do not confuse current per-task controls and provider routing with a finished finance dashboard. Forge will label those controls by availability as they move from policy contracts into the product surface.
A practical low-waste pattern
Start with one lead and one narrowly scoped researcher or implementer. Gather current source truth, write one failing test at the public seam, make the smallest coherent change, run the targeted deterministic gates, and request one independent final review. Expand the team only when the remaining work is truly parallel.
That pattern cuts usage because it cuts confusion. The team spends fewer turns rediscovering context, fewer reviews on unstable code, and fewer expensive model calls answering questions the repository can answer directly.