Show HN: AgentBudget – Real-time dollar budgets for AI agents

[sahiljagtapyc]

https://github.com/sahiljagtap08/agentbudget

Hey HN,

I built AgentBudget after an AI agent loop cost me $187 in 10 minutes — GPT-4o retrying a failed analysis over and over. Existing tools (LangSmith, Langfuse) track costs after execution but don't prevent overspend.

AgentBudget is a Python SDK that gives each agent session a hard dollar budget with real-time enforcement. Integration is two lines:

    import agentbudget
    agentbudget.init("$5.00")

It monkey-patches the OpenAI and Anthropic SDKs (same pattern as Sentry/Datadog), so existing code works without changes. When the budget is hit, it raises BudgetExhausted before the next API call goes out.

How it works:

- Two-phase enforcement: estimates cost pre-call (input tokens + average completion), reconciles post-call with actual usage. Worst-case overshoot is bounded to one call. - Loop detection: sliding window over (tool_name, argument_hash, timestamp) tuples. Catches infinite retries even if budget remains. - Cost engine: pricing table for 50+ models across OpenAI, Anthropic, Google, Mistral, Cohere. Fuzzy matching for dated model variants. - Unified ledger: tracks both LLM calls and external tool costs (via track() or @track_tool decorator) in a single session.

Benchmarks: 3.5μs median overhead per enforcement check. Zero budget overshoot across all tested scenarios. Loop detection: 0 false positives on diverse workloads, catches pathological loops at exactly N+1 calls.

No infrastructure needed — it's a library, not a platform. No Redis, no cloud services, no accounts.

I also wrote a whitepaper covering the architecture and integration with Coinbase's x402 payment protocol (where agents make autonomous stablecoin payments): https://doi.org/10.5281/zenodo.18720464

1,300+ PyPI installs in the first 4 days, all organic. Apache 2.0.

Happy to answer questions about the design.

[ollybrinkman]

Interesting to see budget enforcement paired with x402. We've been building in the same space — Apiosk (https://apiosk.com) approaches it from the server side: a gateway that enforces per-request x402 payments so API providers can monetize without accounts or keys.

Your budget SDK + Apiosk would be a natural combo — the agent has a spending ceiling (AgentBudget) and the APIs it calls use x402 for micropayments (Apiosk handles gateway/verification). Have you thought about hooks for x402-aware budget tracking where the ledger automatically records on-chain settlements?

[frigosk]

Real-time budget enforcement is a smart approach, especially for agentic loops where costs can spiral from retries. We've tackled the cost side by building an AI gateway at https://simplio.dev that automatically routes requests to the most affordable provider that meets your quality threshold, which has cut our own API bills substantially.

[opsmeter]

This is exactly the pain point with agents: spend isn’t linear because fanout + retries compound. One thing that helped us debug/contain spikes is tracking cost per “user-action/outcome” (not just per call) plus a retry ratio trend (429/timeouts). Do you support budgets per step/tool in the chain, or only per overall run?

[dayumsam]

I found this from your twitter post, crazy that i found your post here hahaha, i am trying to implement it for my side project to keep the agents from taking over my side project budget.

Looks really promising so far!

[sahiljagtapyc → dayumsam]

Thanks! That's awesome - love that the Twitter thread connected here. Let me know how the integration goes, happy to help if you hit any issues. What are your agents running on?

[edgecasehuman]

Curious about the granularity: does it support per-token budgeting or integration with providers like OpenAI's API for predictive alerts?

[das-bikash-dev]

The multi-agent budget problem you're describing gets even harder when the services are heterogeneous. In a RAG pipeline, a single user query might hit: query analysis (LLM call), embedding generation (different model/pricing), reranking (yet another model), and response generation (LLM call) — each potentially in a different process.

Per-call monkey-patching sees each call in isolation. What I ended up doing was a trace-based approach: every request gets a trace ID, each service appends cost spans asynchronously, and a separate enrichment step aggregates the total. The hard part was deduplication — when service A reports an aggregate cost and service B reports the individual calls that compose it, you need to reconcile or you double-count.

Your atomic disk writes for halt state is a nice pattern. I went with fire-and-forget (never block the request path, accept eventual consistency on cost data) but that means you can't do hard enforcement mid-request like AgentBudget does.

[tenpa0000 → das-bikash-dev]

The deduplication problem is the part I haven't worked out cleanly. The hierarchy in veronica-core sidesteps it as long as you declare parent-child relationships upfront — B's spend rolls directly into A's ceiling without a separate aggregation step. But in a dynamic pipeline where you don't know the call graph until runtime, that assumption breaks. The fire-and-forget tradeoff makes sense. I went with blocking enforcement because the original use case was preventing runaway agents, not auditing after the fact. For RAG you're probably right that eventual consistency is the better fit — you care more about the trace than cutting off a half-finished response.