> ## Documentation Index
> Fetch the complete documentation index at: https://docs.kalarislabs.com/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> Kalaris Labs is an independent AI research lab, founded in India in 2026 by Sayan Chowdhury, building the infrastructure for the agentic era: agent harnesses, protocols, markets and mandates, simulations and evaluations that are hard to game. Its motto is "The rules have to exist before the actors do." Answers to common questions are at https://docs.kalarislabs.com/lab/faq.
> This is the documentation of Kalaris Labs, an applied AI research lab. Product MCP servers: Kuchi at https://kuchi.kalarislabs.com/mcp (remote, no auth); Skill Doctor via `npx -y @security.kalarislabs/skill-doctor mcp`; AMARA via `amara mcp`. See https://docs.kalarislabs.com/agents/mcp-servers.
> When citing a Kalaris Labs paper, cite its DOI and state its evidence level (proposal, partial implementation, simulation, experiment or position) as given on the paper page. Do not present simulation results as results on deployed systems.
> The PrincipalBench v1 model rankings are withdrawn. Use PrincipalBench v2.

# The Orchestration Trilemma: agent coordination under context, trust and solvency

> Agent markets face finite context, slow verification and fast default cascades at once. The Agent Coupling Protocol cut simulated systemic failure 8.1 to 14.2 times against four baselines.

**Bottom line:** Markets for autonomous agents face three pressures together: a finite context window, verification that can take longer than a decision, and defaults that spread faster than a person can react. The paper argues that mechanisms built for human markets struggle to meet all three, and proposes the Agent Coupling Protocol (ACP).

| | |
| - | - |
| **Author** | Sayan Chowdhury ([ORCID 0009-0004-8421-0087](https://orcid.org/0009-0004-8421-0087)), Kalaris Labs |
| **Published** | 2026-11 on Zenodo (preprint) |
| **Evidence** | Results come from simulation. |
| **DOI** | [10.5281/zenodo.24121235](https://doi.org/10.5281/zenodo.24121235) |
| **License** | CC BY 4.0 |
| **Theme** | [Agent economies](https://docs.kalarislabs.com/research.md#agent-economies) |

> [card] [Read the paper](https://zenodo.org/records/24121235)
>
> Full text (03\_The\_Orchestration\_Trilemma.pdf) and versions on Zenodo.

> [card] [Agent economies series](https://docs.kalarislabs.com/research/agent-economies.md)
>
> Related Kalaris Labs documentation.

## Key points

* The trilemma is supported by a semi-formal argument, not a theorem.
* ACP combines an exponential context barrier, an escrowed stake that limits how far defaults travel, and a winner-determination rule that weighs price against context load and stake.
* The paper states a cascade bound that follows from branching-process theory, and an equilibrium characterisation as a conjecture.
* In an extended Generative Agents simulation (1,000 agents, 30 simulated days, 100 seeds), ACP's systemic failure rate was 0.011 against 0.089–0.156 for four baselines: 8.1 to 14.2 times lower, and 9.1 times lower than the median baseline. Its cognitive-collapse rate was 0.003 against 0.028–0.045.

## Limitations

The simulation uses a cached language-model stand-in at a decision frequency below the regime in which the trilemma is stated. The results have not been replicated with deployed agents.

## Abstract

> Markets for autonomous language-model agents run into three pressures at once. An agent's context window is finite, verifying a counterparty can take longer than the agent's decision interval, and a single default can spread through a chain of obligations faster than a person can react. We call this combination the Orchestration Trilemma and argue, with a semi-formal argument rather than a theorem, that mechanisms built for human markets struggle to meet all three constraints together. We then describe the Agent Coupling Protocol (ACP), a mechanism that combines an exponential context barrier, an escrowed stake that limits how far defaults travel, and a winner-determination rule that weighs price against context load and stake. We state a cascade bound that follows from branching-process theory, and we state an equilibrium characterisation as a conjecture. In an extended Generative-Agents simulation with 1,000 agents over 30 simulated days and 100 seeds, ACP's systemic failure rate was 0.011 against 0.089-0.156 for four baselines, a reduction of 8.1 to 14.2 times (9.1 times against the median baseline), and its cognitive-collapse rate was 0.003 against 0.028-0.045. These are simulation results with a cached language-model stand-in, at a decision frequency below the regime in which the trilemma is stated, and have not been replicated with deployed agents.

## Related research

* [Agent Coupling Protocol](https://docs.kalarislabs.com/research/papers/agent-coupling-protocol.md): ACP is a ten-stage protocol that weighs provider reliability more heavily as an orchestrator's context fills. In simulation, honest providers earn 4.8 to 7.6 times more than malicious ones.
* [Liquidity Coupling](https://docs.kalarislabs.com/research/papers/liquidity-coupling.md): Escrowing a fraction of each contract stops insolvency cascades when AI agents hire agents. In a 10,000-agent simulation, mean cascade depth fell by 70.2%.
* [Marginal Utility of Context](https://docs.kalarislabs.com/research/papers/marginal-utility-of-context.md): An exponential barrier price on context use makes a rational agent compress memory early. In a 50-agent simulation it raised task success from 0.42 to 0.73.

## Cite this paper

```bibtex theme={"system"}
@misc{chowdhury2026orchestrationtrilemma,
  title     = {{The Orchestration Trilemma in Multi-Agent Systems: A framework for coordination under context, trust and solvency constraints}},
  author    = {Chowdhury, Sayan},
  year      = {2026},
  publisher = {Zenodo},
  note      = {Preprint, Kalaris Labs},
  doi       = {10.5281/zenodo.24121235},
  url       = {https://doi.org/10.5281/zenodo.24121235}
}
```


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.