Review: surveys first, then the seminal works, then practice

Step 2, read on 2026-11-02. Questions Q1 to Q4 are in frame; grades in sources.

Shape: review

0. Existing reviews

  • O'Neill 2020 reviews 76 empirical articles on human-autonomy teaming (read in the abstract, secondary). Its variables: agent characteristics, team composition, task, individual differences, training, communication. Grade: supported with caveat.
  • Berretta 2023 is a scoping review and network analysis that derives one definition of human-AI teaming. Grade: supported with caveat.
  • Mosqueira-Rey 2022 and Wu 2022b survey human-in-the-loop for machine learning. Both concern training a model. Grade: convention.
  • Zou 2025 surveys systems where a person gives a language-model agent information, feedback or control, and names fully autonomous agents as unreliable.
  • Vaccaro 2024 is a meta-analysis, placed in against and state of the art.

1. The seminal works

  • Engelbart 1962 defines augmenting human intellect as increasing a person's capability to approach a complex problem. His unit is a person with language, artifacts, methodology and training.
  • Licklider 1960 splits the work: people set the goals, formulate hypotheses, set criteria and evaluate; machines do the routinizable work.
  • Parasuraman 2000 gives four functions to automate (information acquisition, analysis, decision and action selection, action implementation), each across levels from manual to automatic.
  • Dellermann 2019 defines hybrid intelligence as reaching complex goals by combining human and artificial intelligence, with each learning from the other (read in the author copy).
  • Akata 2020 defines hybrid intelligence as combining human and machine intelligence, augmenting and not replacing, and sets four challenges.

2. Teaming and design

  • Seeber 2020 asks what changes when machines are teammates, and gives three design areas: machine artifact, collaboration, institution.
  • Shneiderman 2020 argues for high human control together with high automation.
  • Amershi 2019 gives 18 design guidelines for human-AI interaction, tested with 49 design practitioners on 20 popular AI-infused products.
  • Raisch 2021 argues that augmentation cannot be neatly separated from automation.
  • Hassani 2020 contrasts replacing the human mind with augmenting it, and chooses augmentation.
  • Case 2018 popularises the centaur: people working with machines in a non-zero-sum game.

3. Practice

  • Anthropic 2024 separates workflows, where code orchestrates a model, from agents, where the model directs itself, and says agents may pause for human feedback at checkpoints. It is the owner's account of its own practice, and scores below the threshold.
  • Agentic AI Foundation 2026 is a convention for a file that briefs coding agents; the site claims use by over 60 thousand open-source projects. It is the owner's claim and scores below the threshold.
  • Practice runs ahead of the evidence: both conventions have adoption and no controlled study of their effect.

4. Forward citations, added 2026-11-03

Four works cite the three most-cited sources since 2024. They sharpen three findings.

  • The levels of automation are contested. A 2025 systematic review of levels-of-automation taxonomies finds critiques of misinterpretation and of conceptual confusion between automation and autonomy (Richardson 2025). This qualifies Parasuraman 2000: its levels are a useful vocabulary, not a settled measure.
  • Complementarity is the test, and it is rarely observed. Hemmer 2025 defines the complementary team performance that neither the person nor the model reaches alone, and names information and capability asymmetry as its two sources. González 2026 gives a framework of the conditions for that performance. Both sharpen Vaccaro 2024 and answer Q3 with the conditions, not only the average.
  • Current interaction is narrow. A systematic review of 105 articles on AI-assisted decisions finds that most interactions follow simplistic collaboration paradigms, with little truly interactive support (Gómez 2025, secondary, read in the abstract). This supports the family's rule that the person decides and checks where each claim stands.

What the strands say

Question Answer from this pool
Q1 five families of names: augmentation, symbiosis, hybrid intelligence, teaming, human-in-the-loop
Q2 intelligence augmentation and hybrid intelligence; see position
Q3 not always: see against
Q4 Akata 2020 and Parasuraman 2000: see baseline