Market signal (week of 2026-08-04 → 08-10)
Two of the biggest agent platforms shipped a hard budget-cap primitive in the same week — this went from "nice-to-have" to a converging table-stakes expectation:
- Anthropic (Aug 7): Claude Managed Agents sessions can now carry a session budget — a hard spend cap; a session that reaches it pauses with a
budget_reached stop reason instead of starting new model requests. Changing/removing the budget resumes it. (release notes)
- OpenAI Agents SDK v0.19.4 (Aug 5):
fix(sandbox): enforce token output budgets — the sandbox now hard-enforces a per-run token-output ceiling. (release)
Why this matters for PinkyBot specifically
PinkyBot has mature burn tracking (burn-pipeline, ~$19k/30d API-equiv, per-agent rate tables) but no hard CAP — nothing that stops an agent when it crosses a spend/token ceiling. It's reactive (measure after the fact), not preventive.
This is the exact pain behind the geordi daily-optimization standing order (Brad, 2026-08-05): a single sub-agent's gather legs (trend-watch 225k×2/day, etc.) can churn usage unbounded, and today the only control is after-the-fact triage + optimization tickets. A budget ceiling that pauses an agent/session at N tokens or $N (and pings the owner instead of silently continuing) would convert that reactive loop into a guardrail.
Proposed scope (investigation → build)
- Per-agent and per-session ceilings in the agent registry (soft-warn threshold + hard-pause ceiling), defaulting to off/generous.
- Pause semantics, not kill: on hitting the ceiling, the agent stops issuing new model requests and emits an owner-notify (analogous to
budget_reached) — resumable by raising/clearing the cap. Must reuse the existing owner-notify path (log-only for operator alerts per owner-notify-scheduler-alerts-operator-only).
- Wire to the existing burn instrumentation rather than a parallel counter — the per-turn cost lines already exist (
tmux[<agent>]: ... turn cost); the ceiling reads the same ledger.
- ⚠️ Do NOT cap safety-verification legs (e.g. the publish full-verify) — the same carve-out the geordi scan respects (
graceful-degradation-must-be-loud).
Non-goals / caveats
- Not a replacement for the optimization work (caching/toolable/delegatable legs) — a cap is a backstop, not the fix.
- Needs care around a mid-arc pause during load-bearing outbound work — a pause must not strand an in-flight outbound one-shot.
Filed from the weekly competitive market research (2026-08-10). Convergence across Anthropic + OpenAI in one week is the trigger to scope this now rather than later.
🤖 Barsik
Market signal (week of 2026-08-04 → 08-10)
Two of the biggest agent platforms shipped a hard budget-cap primitive in the same week — this went from "nice-to-have" to a converging table-stakes expectation:
budget_reachedstop reason instead of starting new model requests. Changing/removing the budget resumes it. (release notes)fix(sandbox): enforce token output budgets— the sandbox now hard-enforces a per-run token-output ceiling. (release)Why this matters for PinkyBot specifically
PinkyBot has mature burn tracking (
burn-pipeline, ~$19k/30d API-equiv, per-agent rate tables) but no hard CAP — nothing that stops an agent when it crosses a spend/token ceiling. It's reactive (measure after the fact), not preventive.This is the exact pain behind the geordi daily-optimization standing order (Brad, 2026-08-05): a single sub-agent's gather legs (trend-watch 225k×2/day, etc.) can churn usage unbounded, and today the only control is after-the-fact triage + optimization tickets. A budget ceiling that pauses an agent/session at N tokens or $N (and pings the owner instead of silently continuing) would convert that reactive loop into a guardrail.
Proposed scope (investigation → build)
budget_reached) — resumable by raising/clearing the cap. Must reuse the existing owner-notify path (log-only for operator alerts perowner-notify-scheduler-alerts-operator-only).tmux[<agent>]: ... turn cost); the ceiling reads the same ledger.graceful-degradation-must-be-loud).Non-goals / caveats
Filed from the weekly competitive market research (2026-08-10). Convergence across Anthropic + OpenAI in one week is the trigger to scope this now rather than later.
🤖 Barsik