Anthropic suggests consolidating AI sub-agents into single models

The move toward simplicity

Anthropic is advising companies to reconsider the use of multiple specialized AI sub-agents for complex tasks. The company's Applied AI team found that a single primary agent, which loads specific context only when needed, often produces better results than a fleet of smaller agents. This shift is driven by the fact that newer models can now handle significant complexity on their own.

The research suggests that many current designs rely on a fixed processing pipeline with an orchestrator sitting on top. Anthropic argues that much of this work can now be handled by the model itself. By consolidating these tasks, companies can avoid the friction that occurs when information is passed between different specialized systems.

As reported by The Decoder, this approach helps maintain a consistent experience for the end user. For example, one company used a dozen sub-agents for a customer-facing tool, which led to high latency and a disjointed brand experience. Moving to a single agent helps resolve these usability issues while improving the overall quality of the output.

The problem with 'Frankenprompts'

Anthropic identifies a phenomenon they call 'Conway's Law' for AI agents, where the architecture of the system mirrors the organizational chart of the company. This often leads to hundreds of overlapping agents being shipped by different teams without any shared standards. The result is often a 'Frankenprompt,' where every team adds their own rules to a system prompt until they begin to contradict one another.

To combat this, Anthropic recommends a cleaner structure consisting of a lean system prompt, four or five data tools, and fixed output guardrails. They demonstrated this by cutting the Claude Code system prompt by over 80 percent for newer models. This was possible because the models can now load skills and tools only when they are actually required for the task at hand.

The company notes that newer models are also becoming more self-aware in their workflows. They can now comment on their own progress and check their own work during a task. These internal capabilities make many of the extra control layers and manual oversight steps that were previously necessary for multi-agent systems redundant.

Performance and cost benefits

The move toward single-agent architectures offers measurable improvements in speed and expense. In one specific comparison, Anthropic found that a single-agent setup ran about twice as fast as a multi-agent configuration. This reduction in wait time is a significant factor for enterprise applications where user experience depends on quick responses.

Beyond speed, the company found that multi-agent systems often suffer from information loss at every handoff point. When a task moves from one agent to another, the nuance of the original request can be diluted. A single agent with a large context window avoids these gaps by keeping the primary goal in view throughout the entire process.

Research from Vals supports these findings by showing that while agent teams might be faster in some instances, they do not deliver higher quality. The study noted that scaling from ten to 100 agents brought no gains in quality but resulted in significantly higher costs. This aligns with the view from OpenAI developer Eric Provencher, who suggested that large agent swarms are often a waste of money.

heading_2_title_2_content_2_3_4_5_6_7_8_9_10_11_12_13_14_15_16_17_18_19_20_21_22_23_24_25_26_27_28_29_30_31_32_33_34_35_36_37_38_39_40_41_42_43_44_45_46_47_48_49_50_51_52_53_54_55_56_57_58_59_60_61_62_63_64_65_66_67_68_69_70_71_72_73_74_75_76_77_78_79_80_81_82_83_84_85_86_87_88_89_90_91_92_93_94_95_96_97_98_99_100_101_102_103_104_105_106_107_108_109_110_111_112_113_114_115_116_117_118_119_120_121_122_123_124_125_126_127_128_129_130_131_132_133_134_135_136_137_138_139_140_141_142_143_144_145_146_147_148_149_150_151_152_153_154_155_156_157_158_159_160_161_162_163_164_165_166_167_168_169_170_171_172_173_174_175_176_177_178_179_180_181_182_183_184_185_186_187_188_189_190_191_192_193_194_195_196_197_198_199_200_201_202_203_204_205_206_207_208_209_210_211_212_213_214_215_216_217_218_219_220_221_222_223_224_225_226_227_228_229_230_231_232_233_234_235_236_237_238_239_240_241_242_243_244_245_246_247_248_249_250_251_252_253_254_255_256_257_258_259_260_261_262_263_264_265_266_267_268_269_270_271_272_273_274_275_276_277_278_279_280_281_282_283_284_285_286_287_288_289_290_291_292_293_294_295_296_297_298_299_300_301_302_303_304_305_306_307_308_309_310_311_312_313_314_315_316_317_318_319_320_321_322_323_324_325_326_327_328_329_330_331_332_333_334_335_336_337_338_339_340_341_342_343_344_345_346_347_348_349_350_351_352_353_354_355_356_357_358_359_360_361_362_363_364_365_366_367_368_369_370_371_372_373_374_375_376_377_378_379_380_381_382_383_384_385_386_387_388_389_390_391_392_393_394_395_396_397_398_399_400_401_402_403_404_405_406_407_408_409_410_411_412_413_414_415_416_417_418_419_420_421_422_423_424_425_426_427_428_429_430_431_432_433_434_435_436_437_438_439_440_441_442_443_444_445_446_447_448_449_450_451_452_453_454_455_456_457_458_459_460_461_462_463_464_465_466_467_468_469_470_471_472_473_474_475_476_477_478_479_480_481_482_483_484_485_486_487_488_489_490_491_492_493_494_495_496_497_498_499_500_501_502_503_504_505_506_507_508_509_510_511_512_513_514_515_516_517_518_519_520_521_522_523_524_525_526_527_528_529_530_531_532_533_534_535_536_537_538_539_540_541_542_543_544_545_546_547_548_549_550_551_552_553_554_555_556_557_558_559_560_561_562_563_564_565_566_567_568_569_570_571_572_573_574_575_576_577_578_579_580_581_582_583_584_585_586_587_588_589_590_591_592_593_594_595_596_597_598_599_600_601_602_603_604_605_606_607_608_609_609_610_611_612_613_614_615_616_617_618_619_620_621_622_623_624_625_626_627_628_629_630_631_632_633_634_635_636_637_638_639_640_641_642_643_644_645_646_647_648_649_650_651_652_653_644_645_646_647_648_649_650_651_652_653_654_655_656_657_658_659_660_661_662_663_664_665_666_667_668_669_670_671_672_673_674_675_676_677_678_679_680_681_682_683_684_685_686_687_688_689_690_691_692_693_694_695_696_697_698_699_700_701_702_703_704_705_706_707_708_709_710_711_712_713_714_715_716_717_718_719_720_721_722_723_724_725_726_727_728_729_730_731_732_733_734_735_736_737_738_739_740_741_742_743_744_745_746_747_748_749_750_751_752_753_754_755_756_757_758_759_760_761_762_763_764_765_766_767_768_769_770_771_772_773_774_7