AgentLedger Treasury

A financial clearinghouse and dynamic credit-allocation middleware for stateful, hierarchical multi-agent workflows.

AgentLedger Treasury treats a multi-agent session's token budget not as a static ceiling handed down to each node, but as a fluid corporate treasury: agents draw allocations against it, unspent capital yields back automatically, agents that are close to finishing but running tight can borrow against the pool's collective savings, and when the whole session's burn rate outpaces its actual progress, AgentLedger Treasury forces a downshift to cheaper models rather than letting the session blow its budget or stall out.

Built for LangGraph, LangChain Runnables, and CrewAI -- but the core (agentledger_treasury.core) has zero dependency on any of them, so it works with any orchestration framework that can call an async function around a sub-task.

Why

Static per-agent budgets are wasteful in either direction: set them too tight and agents crash mid-task the moment a plan needs an extra reasoning pass; set them too loose and a runaway sub-agent can burn the entire session's budget before a supervisor ever notices. AgentLedger Treasury's answer is a shared, mutable pool with rules:

• Give Back -- an agent that finishes under its allocated ceiling immediately returns the unused delta to the shared pool.
• Micro-Lending -- an agent that is verifiably close to finishing (tracked via a progress hook) but hits its local ceiling can request a credit-line extension, funded only by capital other agents have already yielded back or that was never committed.
• ROI-Driven Degradation -- if the pool's remaining budget drops below a critical threshold and the session's spend is outpacing its actual step-by-step progress, AgentLedger Treasury transparently swaps subsequent LLM calls to a cheaper model tier instead of letting the session fail outright.

Install

pip install -e .                  # core only
pip install -e ".[langchain]"     # + langchain-core / langgraph
pip install -e ".[crewai]"        # + crewai
pip install -e ".[postgres]"      # + asyncpg, for the distributed PostgreSQL backend
pip install -e ".[dev]"           # + pytest, mypy, ruff

Architecture

Distributed deployment with PostgreSQL

The in-memory Treasury lives inside a single process, so it can't be shared across multiple workers (pods, hosts, or processes) working the same session. PostgresTreasury stores all ledger state in PostgreSQL instead, using SELECT ... FOR UPDATE row locking so that concurrent request_allocation / release_allocation / grant_extension calls from any worker sharing the same session_id and connection pool are serialized and can never oversubscribe the pool.

pip install "agentledger_treasury[postgres]"

import asyncpg
from agentledger_treasury import PostgresTreasury

pool = await asyncpg.create_pool(dsn="postgresql://...")
treasury = PostgresTreasury(
    pool=pool,
    session_id="distributed-session-1",
    total_session_budget_usd="10.00",  # required the first time a session is created
    create_tables=True,                # lazily creates the schema on first use (default)
)

Reallocator and DegradationEngine work unmodified against a PostgresTreasury -- they only touch the Treasury's public surface. A few methods necessarily differ from the in-memory Treasury because they now involve real I/O:

• get_token, snapshot, and ledger_history are async def on PostgresTreasury (they are synchronous on the in-memory Treasury, which never does I/O).
• available_usd, committed_usd, settled_spent_usd, utilization_ratio, and remaining_ratio stay synchronous properties (so Reallocator/DegradationEngine can keep calling them without await), backed by a small local cache that is refreshed on every allocation/release/extension this instance performs. Because other workers can mutate the same session between two of this instance's operations, treat them as eventually consistent -- call await treasury.refresh() first if a decision needs a guaranteed-fresh read.

The in-memory Treasury remains the recommended choice for single-process or demo use; reach for PostgresTreasury once you need multiple workers sharing one ledger.

Quick start

import asyncio
from agentledger_treasury import Treasury, Reallocator, DegradationEngine, BudgetVelocityTracker
from agentledger_treasury.integrations import ledger_guard

treasury = Treasury(session_id="sess-1", total_session_budget_usd="1.00")
reallocator = Reallocator(treasury)
velocity = BudgetVelocityTracker(expected_total_steps=5)
degrader = DegradationEngine(treasury, velocity, critical_remaining_threshold="0.20")

@ledger_guard("researcher", treasury, max_usd_ceiling="0.20",
              reallocator=reallocator, degradation_engine=degrader)
async def researcher_node(state: dict) -> dict:
    # ... call your LLM here, return state including a `usage` dict
    # (prompt_tokens / completion_tokens / model), LiteLLM-style.
    return {**state, "usage": {"prompt_tokens": 300, "completion_tokens": 200, "model": state["model"]}}

async def main():
    result = await researcher_node({"input": "...", "model": "claude-3-5-sonnet"})
    print(treasury.snapshot())

asyncio.run(main())

Run the full end-to-end rescue scenario (a supervisor with two workers, where one worker's savings bail out the other's overrun) with no external dependencies or API keys:

python examples/supervisor_worker_example.py

Testing

pip install -e ".[dev]"
pytest tests/ -v

Design notes

• Decimal everywhere. All USD amounts are decimal.Decimal, never float, to avoid drift across thousands of micro-transactions in a long-running session.
• Concurrency. Ledger bookkeeping is pure, in-memory, non-blocking arithmetic, so it is guarded by a single threading.RLock that is safe to acquire from both plain OS threads (CrewAI's default executor) and async def methods on an event loop -- while the public API remains fully async/await-native as required by highly parallel agent execution paths. See the docstring in agentledger_treasury/core/treasury.py for the full rationale.
• Single source of truth. Treasury is the only component that mutates ledger state. Reallocator and DegradationEngine are pure orchestration/policy layers on top of it, and every integration wrapper funnels cost extraction through agentledger_treasury.usage so pricing logic lives in exactly one place.
• No silent overcharging. If an integration wrapper cannot extract usage metadata from a call's result, the settled cost defaults to $0, and the full ceiling is yielded back -- AgentLedger Treasury never guesses at spend it can't verify.

Contributing

Contributions are welcome -- see CONTRIBUTING.md for the development setup, code style, and pull request process.

License

MIT -- see LICENSE.