Each step does one thing with a narrowed highly specific context, oblivious to the general noise "outside" in the broader context.
Includes :
- rule based actions: sequences, loops, conditions, grouping, map reduce, filtering, recursion and data merging
- neuro symbolic reasoning, contextual
math and statistics, semantic searching
- response reassessment, pre-response sequences, post response actions (like secure permission consolidation, complex app access)
- verifications: checking logic, coherence, consistence, comprehensiveness, authenticity, acceptance,
- rating: confidence, guesswork, knowledge gap, user alignment
- text manipulation, data manipulation, source retrieval,
- on the fly coding, human in the loop actions
- tool improvement (OCR, scraping, speech-reco etc) especially by exposing under the hood middle or raw layers
- conversation improvement: active listening, knowledge probing, topic mapping, staying-on-track elements
The step content could be pre-conceived or constructed with code on the fly, with detailed instructions for rule based, ai based, or hybrid flows to construct the desired result.
This new type of tool could be supplied by the LLM companies, or if they don't bite in, could be supplied through a community and used some free some for pay (ie improved contextual speech recog or ocr) in ai snippets