Resource-bounded agent contracts
Resource-bounded agent contracts are the Ye-Tan Agent Contracts method for specifying a delegated agent run before activation. Its formal contract combines input and output specifications, allowed skills, multi-dimensional resource budgets, temporal limits, success criteria and termination conditions. A lifecycle records activation and a terminal outcome, while parent-child conservation rules constrain how an orchestrator allocates budgets to delegated agents. The qualified name separates this resource-governance method from other, behavior-oriented uses of `agent contract`.
Origin and context
Ye and Tan submitted the framework to arXiv in January 2026 and presented it in the organizations-and-governance session of the COINE 2026 workshop co-located with AAMAS. A Python package followed and reached version 0.5.0 in August 2026. Separate papers then treated the framework as resource governance, contrasted it with behavioral contracts and oversight allocation, and directly tested its runtime cap behavior.
Why it matters
Ordinary per-call settings do not express a whole workflow's combined token, cost, tool-call, iteration and time envelope. A contract can put those dimensions, acceptable output and stop conditions in one inspectable object, then propagate smaller allocations into sub-agents. That makes intended limits and allocation mistakes easier to review. It does not make model behavior deterministic or guarantee that actual provider charges and side effects remain below every declared number.
Example
A coordinator receives a research task with a total token, API-call, tool and duration budget. Before spawning researcher and writer agents, it reserves child allocations whose sum fits the parent contract. A wrapper checks the remaining allowance before each mediated call, records returned usage afterwards and prevents later calls when a limit is reached. If one LLM call itself overshoots, the excess can still occur: current APIs generally reveal final usage only when that call completes.
How it differs
Agent Behavioral Contracts / ABC
Agent Behavioral Contracts specify preconditions, invariants, governance and recovery for behavior over time. Their own paper calls the Ye-Tan framework complementary resource governance, so neither tuple nor evidence should be merged into the other.
AgentSpec runtime-enforcement DSL
AgentSpec is a domain-specific language whose trigger, predicate and enforcement rules intercept planned actions. A resource-bounded contract can use such a policy mechanism, but its defining concern is the run-level resource, time, output and delegation envelope.
Reasoning Effort and Thinking Budget
Reasoning effort or a thinking budget controls inference within a model call. A resource-bounded contract spans multiple calls, tools and agents and must account for provider-reported usage after execution; the two controls can be layered.
Token Cost Attribution
Token cost attribution assigns observed or billed usage to an owner or workload. Contracts declare and enforce a budget policy; their audit records may feed attribution, but neither accurate allocation nor invoice reconciliation follows from a contract declaration alone.
Maturity and evidence
Maturity is rated 3. The method has a precise formal definition, an official workshop presentation, an actively released package, independent same-sense research use and an independent head-to-head experiment. It remains below 4 because no neutral standard or broad multi-organization production adoption was located, APIs have changed across pre-1.0 releases, and project-owned results do not establish general effectiveness.
Limits and open questions
Enforcement is only as complete as mediation and measurement. A single LLM call can exceed a token or cost limit before usage is visible, unwrapped call sites can bypass checks, and the current project does not enforce iteration limits uniformly across integrations. Success predicates and audit events can be incomplete, provider accounting can drift, and external actions may already be irreversible. This software formalism is not a legal contract, compliance certificate, safety proof or guarantee of task quality.
Related terms
References
- Agent Contracts: A Formal Framework for Resource-Bounded Autonomous AI SystemsQing Ye and Jing Tan / arXiv · 2026-01-13 · class A
- COINE 2026 Technical ProgrammeCOINE Workshop · 2026 · class A
- flyersworder/agent-contractsQing Ye · 2026-08-30 · class A
- ai-agent-contracts 0.5.0Python Package Index · 2026-08-22 · class A
- Token Budgets: An Empirical Catalog of 63 LLM-Agent Budget-Overrun Incidents, with an Affine-Typed Rust Mitigation as a Case StudySajjad Khan / arXiv · 2026-06-02 · class B
- Agent Behavioral Contracts: Formal Specification and Runtime Enforcement for Reliable Autonomous AI AgentsVarun Pratap Bhardwaj / arXiv · 2026-02-25 · class B
- Minimal Oversight: Uncertainty-Aware Governance for Delegated AI SystemsCarlos R. B. Azevedo / arXiv · 2026-06-04 · class B
- Agent Contracts: A Framework for Reliable AI SystemsRelari · 2025-04-30 · class A
- AgentSpec: Customizable Runtime Enforcement for Safe and Reliable LLM AgentsHaoyu Wang, Christopher M. Poskitt and Jun Sun / arXiv · 2025-03-24 · class A
- Agent Operating Systems (Agent-OS): A Blueprint Architecture for Real-Time, Secure, and Scalable AI AgentsAnis Koubaa / Preprints.org · 2025-09-01 · class B
Last updated: 2026-09-07
This term is also covered in the Skills Atlas as multi agent systems skill.
This term is also covered in the Skills Atlas as ai finops skill.