Large-scale real-time high-performance data acquisition computing systems often require to be fault-tolerant and adaptive to changes. We consider a multi-agent system based approach to achieve these goals. This research is a part of ongoing research efforts to build a triggering and data acquisition system (known as BTeV) for particle-accelerator-based high energy physics experiments at Fermi National Laboratory. The envisioned hardware consists of pixel detectors and readout sensors embedded in the accelerator, which are connected to specialized FPGAs (field-programmable gate arrays). The FPGAs are connected to approximately 2,500 digital signal processors (DSPs). After initial filtering and processing of data by the DSPs, a farm of approximately 2,500 Linux computers are responsible for post-processing a large amount of high speed data input. To support the adaptive fault-tolerance feature, we introduce the notion of very lightweight agents (VLAs), which are designed to be adaptive but small in footprint and extremely efficient. Each digital signal processor will run a very lightweight agent along with a physics application program that collects and processes input data from the corresponding FPGA. Since VLAs can be proactive or reactive, Brooks' subsumption architecture is a good basis for the design. In this paper we present several necessary changes in the original subsumption architecture to better serve the BTeV architecture.