BreakerboxHow it compares
An honest map of the agent-spend landscape — AgentBudget, Alephant, Runyard, LiteLLM, and the observability tools — and the five pieces no single player has yet assembled.
The idea of a "circuit breaker for agents" is in the air. Several tools do a piece of it well; none has assembled the whole thing. This page names them honestly and says where Breakerbox differs — and where it complements rather than competes.
The players
| Player | What they do well | The gap |
|---|---|---|
| AgentBudget | One-line session dollar limit, auto-tracking, circuit breaking, nested budgets, multi-provider. Owns the one-liner. | Session-scoped wrapper. No control plane, no multi-tenant ledger, no reconciliation against billing truth, no policy language, and no framework-native trip points inside a run. |
| Alephant | A proxy with a budget circuit breaker (alert 70% / throttle 90% / kill 100%), thrashing detection, per-agent attribution, efficiency grading. Good ideas on waste-grading. | Requires routing all traffic through their proxy, and enforces at proxy granularity only — it cannot degrade inside a workflow (swap the model at node 7, narrow tools mid-run). |
| Runyard | Swarm / agent / task layered token budgets with breakers. | Bound to their own swarm runtime — not a neutral layer for LangGraph, Claude Code, or a custom stack. |
| LiteLLM & gateways | Key / user / team max_budget on the proxy. | A cautionary tale as much as a competitor: the public tracker shows budget limiters instantiated but never registered (#27381), team-key bypasses (#12905), and regressions where spend continues past the cap (#26672). Enforcement bolted onto a router as feature #47 doesn't get engineered to be trustworthy. |
| Langfuse · Helicone · LangSmith · Datadog | Best-in-class spend observability. | Read-only by architecture — they tell you what you spent. Partners, not competitors: Breakerbox ingests from and exports to them. |
| Knowledge / retrieval layers | Upstream prevention: better retrieval → shorter loops. | Shifts the median of the cost distribution; cannot bound its tail. Complementary — the mature stack is prevent + observe + enforce, and Breakerbox is layer three. |
| Anthropic / OpenAI native caps | Org-level hard caps. | Blunt instruments: no per-workflow granularity, no degradation, single-provider. Breakerbox builds on their surfaces (Claude Code hooks, OTel), never against them. |
The unclaimed square
Enforcement isn't one feature — it's five, and no player above has more than one of them. The product is the combination:
- Framework-native enforcement — inside the loop (LangGraph node boundaries, Claude Code hooks, 1.x middleware), not just in front of it.
- Degradation ladders — swap to a cheaper model and narrow tools at 80%, graceful-stop with partial results at 100%, hard-kill only as the last rung. Not a binary kill switch. (shipped)
- A reconciled, multi-tenant ledger — a fast local counter for the hot path, trued up against OTel / billing ground truth. Enforcement you can trust with money. (roadmap)
- Policies in plain language — the budget owner declares the rule; the system compiles it. Owned by the budget holder, not the YAML holder. (roadmap)
- Shadow mode & audit — log what would trip before you enforce anything; every trip explains its ledger trail. Trustworthy enough for production. (shipped)
Breakerbox leads with #1, #2, and #5 today — the pieces you can't build in an afternoon — and is building toward the control plane (#3, #4). See the roadmap.
The cap is the hook, not the moat. The plain per-run budget cap is table stakes — a small side project can reinvent it in ~400 lines. Lead with the rest: the provable worst-case ceiling before the run, semantic loop detection, the degradation ladder, hierarchical escrow, and the security suite.