Skip to main content

agent-budget: sub-agents spend money their parent no longer has

PyPI Python versions CI Typed: mypy --strict License: MIT

When an orchestrator hands work to several sub-agents at once, each one checks the same "remaining" figure, none of them sees the others, and together they spend a multiple of what was left. The protocols they talk over have no place to say how much a delegated task may spend. The open A2A proposal for one puts it plainly: "There's no way to tell a downstream agent 'you have $0.50 remaining.'" (a2aproject/A2A #2121). agent-budget reserves each sub-agent's slice from the parent before it starts, books every amount on the parent as it happens, and returns the unused part when it finishes.

pip install llm-agent-budget            # one dependency: llm-prices (no dependencies of its own)
pip install "llm-agent-budget[a2a]"     # plus the a2a-python plugins

Fail closed, on purpose

A soft limit records the overspend after it has happened. Since 0.2 agent-budget refuses every call it cannot price exactly:

What other trackers do What agent-budget does Loud opt-out (warns on every call)
Unknown model priced at $0 or a $1/$3 default UnknownModel, call refused Budget(..., allow_unpriced=True)
Cached tokens billed as plain input CachePriceMissing unless the cache rate is listed allow_unpriced=True
Price table pasted into code, never updated StalePrice once a price is 90 days old Pricer(allow_stale=True), max_price_age_days=
Check spend, make the call, book it after refused before the call is made Budget(..., unsafe_soft_limit=True)
A budget with no cap Budget(None) raises none
from agent_budget import Budget

b = Budget(0.10)
hold = b.reserve_for("gpt-4o", {"prompt_tokens": 8_000, "completion_tokens": 1_000})  # $0.03 held
# ... the call, with any client ...
hold.settle_usage("gpt-4o", response.usage)   # OpenAI, Anthropic or Gemini usage, cache tokens included
b.reserve_for("gtp-4o", ...)                  # UnknownModel: refused, not booked at $0

Prices come from llm-prices: every row cites the provider's page and the date it was checked. The opt-outs issue an UnsafeBudgetWarning on every call they let through (never deduplicated), and log it to the agent_budget logger.

How much would a soft limit have let through?

agent-budget audit calls.jsonl --cap 1      # or: python -m agent_budget audit ...

One JSON object per line: model, the provider's usage block, and optionally booked_usd (what your tracker booked). On twelve gpt-5.5 calls booked at a $1/$3 default (the PraisonAI pattern):

Soft limit (checks spend before each call, books after, $0 for unpriced models):
  stopped the run at call 10, after letting 9 calls through
  tracker booked $1.0350; those calls really cost $5.8500
  MONEY PAST THE CAP: $4.8500 (5.85x the cap); real spend crossed $1.00 at call 2, then 7 more calls ran

Hard cap, fail closed (agent-budget default):
  admitted 1 calls, spent $0.6500 of $1.00

Exit status 2 when money got past the cap (a CI gate), --json for machines. Same report as the Rust crate's llm-budget audit.

Child scopes

from agent_budget import Budget

orchestrator = Budget(1.00, name="orchestrator")

with orchestrator.child("researcher", usd=0.60) as research:
    hold = research.reserve(0.06)        # before the model call; refused if it does not fit
    # ... call the model ...
    hold.settle(0.052)                   # what it really cost

    with research.child("fetcher", fraction=0.5) as fetch:   # half of what research has left
        fetch.reserve(0.01).settle(0.01)

print(orchestrator.spent_usd)            # 0.062: every level, as it happens
print(orchestrator.remaining_usd)        # 0.938: the unused slice came back on exit
  • A slice that does not fit raises BudgetExhausted before the sub-agent spends anything.
  • A closed scope (and everything below it) raises ScopeClosed; calls already in flight still settle.
  • kill() stops a scope and its whole subtree. max_depth (default 8) raises DepthExceeded.
  • own_cost_usd + child_cost_usd == spent_usd on every scope.
  • Thread-safe (one lock per tree) and safe in asyncio tasks; current_budget() is the innermost scope entered with with in the current thread or task. One budget can be entered by many threads and tasks at once.

Across agents: A2A, W3C baggage, MCP

BudgetContext is the budget a caller hands down and the cost a callee reports back, in the shape of the A2A budget extension draft (budget{cost, tokens, deadline}, incurred{cost, aggregatedChildCost, tokens, delegations}), plus a reservation{id, amount, currency, expires} for the slice the caller reserved.

from agent_budget import Budget, outbound, to_a2a_metadata, from_a2a_metadata, to_baggage

with Budget(1.00).child("remote", usd=0.30) as slice_:
    ctx = outbound(slice_)
    metadata = to_a2a_metadata(ctx)      # {"https://a2a-protocol.org/extensions/budget/v1": {...}}
    header = to_baggage(ctx)             # "llm.budget.cost.amount=0.3,llm.budget.cost.currency=USD,..."
  • A2A: the whole context under the extension URI in request, message or task metadata.
  • W3C baggage: llm.budget.* members, so anything that already propagates OpenTelemetry baggage carries the budget.
  • MCP: MCP reserves params._meta.baggage for W3C baggage (SEP-414); inject_mcp_meta / extract_mcp_meta put the same members there.

Money is a decimal string on the wire, exact to the nano-dollar. Fields nobody measured are left out, never written as zero.

a2a-python plugins

from a2a import plugins
from agent_budget.a2a_plugin import BudgetClientPlugin, BudgetExecutionPlugin

# Caller: requests made inside a scope carry its budget; replies are booked on it.
plugins.set_global(plugins.sequential(BudgetClientPlugin()))
with orchestrator.child("remote-researcher", usd=0.30):
    async for event in client.send_message(request):
        ...

# Callee: each execution runs inside a Budget of what the caller allowed,
# and every task update reports the cumulative cost back.
handler = DefaultRequestHandlerV2(..., plugins=plugins.sequential(BudgetExecutionPlugin(max_usd=5.00)))
# inside AgentExecutor.execute: current_budget().reserve(...).settle(...)

Strands Agents

from strands import Agent
from agent_budget.strands import Price, SpendCeiling

agent = Agent(tools=[researcher], plugins=[SpendCeiling(usd=0.50, price=Price(input=3.00, output=15.00))])

A Strands plugin (pip install "llm-agent-budget[strands]"): each invocation can spend at most usd, every model call reserves its worst case before it is sent, and an agent used as a tool gets its slice from the caller's budget. See docs/strands.md.

Same format as the Rust crate

llm-budget 0.8 (Rust) has the same scopes and the same contexts. Both test suites read one set of golden fixtures (tests/fixtures/budget_context/), so a Rust orchestrator and a Python agent agree to the last digit.

More: docs/guide.md. Adding it to another project (dependencies, typing, stability, the five-line integration): docs/integrating.md.

MIT license.

Metadata

Release files for llm-agent-budget 0.2.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for llm-agent-budget 0.2.0
File Size Uploaded Provenance Actions
llm_agent_budget-0.2.0.tar.gz 69.8 kB With provenance Details for llm_agent_budget-0.2.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for llm-agent-budget 0.2.0
File Interpreter ABI Platform Actions
llm_agent_budget-0.2.0-py3-none-any.whl Python 3 none any Details for llm_agent_budget-0.2.0-py3-none-any.whl

Total release size: 112.6 kB

Release files / llm_agent_budget-0.2.0.tar.gz

Download URL llm_agent_budget-0.2.0.tar.gz
Size 69.8 kB
Tags Source
SHA-256 checksum
How to use checksums
0317c89acc7315a8656819448ae0afd33fe95e34dfb64c7fe80059c2dfe203c4
BLAKE2b-256 checksum
How to use checksums
b05ebfdd8ce4514c7b9975d86a5ad6be702549f9dbb37852cb1f2387b210d45d
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.

Transparency log

Release files / llm_agent_budget-0.2.0-py3-none-any.whl

Download URL llm_agent_budget-0.2.0-py3-none-any.whl
Size 42.8 kB
Tags Python 3
SHA-256 checksum
How to use checksums
b7ee4022d647957d5cd442df568980c32434bd39861bd13957145c36b8a040b0
BLAKE2b-256 checksum
How to use checksums
59896c05e1ea68d96e71cc07ddab031a17482e20b4b9ab355eb2f4a5862ba8b3
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.

Transparency log

Security for llm-agent-budget 0.2.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for llm_agent_budget-0.2.0.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.2.0 This release

2 release files Full provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page