{
  "id": "agential-turn-in-human-llm-activity",
  "title": "Agential Turn in Human-LLM Activity",
  "author": {
    "name": "Dr. Todd J.B. Blayone",
    "cite_name": "Blayone, T. J. B.",
    "orcid": "0000-0001-6965-7033",
    "profile_url": "../profiles/profile-todd.html"
  },
  "date": "2025-11-25",
  "url": "https://scholarflow.ca/essays/agential-turn-in-human-llm-activity.html",
  "summary": "Chat was only the first public form of human-LLM activity. The real agential turn begins when models, tools, files, and people form durable systems capable of acting beyond a single conversation.",
  "description": "Human-LLM activity beyond chat: how tool-mediated, durable AI systems expand agency, coordination, and social transformation.",
  "tags": [
    "human-AI systems",
    "activity theory",
    "AI empowerment",
    "civic technology",
    "distributed agency"
  ],
  "source_type": "ScholarFlow essay",
  "license": "CC BY-NC 4.0",
  "word_count": 1265,
  "reading_time_minutes": 6,
  "citations": {
    "apa": "Blayone, T. J. B. (2025, November 25). Agential Turn in Human-LLM Activity. ScholarFlow Research. https://scholarflow.ca/essays/agential-turn-in-human-llm-activity.html",
    "bibtex": "@online{agentialturninhumanllmactivity2025,\n  title = {Agential Turn in Human-LLM Activity},\n  author = {Blayone, T. J. B.},\n  year = {2025},\n  month = {November},\n  url = {https://scholarflow.ca/essays/agential-turn-in-human-llm-activity.html},\n  publisher = {ScholarFlow Research},\n  note = {ScholarFlow essay},\n  urldate = {2025-11-25}\n}",
    "ris": "TY  - ELEC\nTI  - Agential Turn in Human-LLM Activity\nAU  - Blayone, T. J. B.\nPY  - 2025\nDA  - 2025-11-25\nPB  - ScholarFlow Research\nUR  - https://scholarflow.ca/essays/agential-turn-in-human-llm-activity.html\nER  -\n"
  },
  "llm_markdown": "---\ntitle: Agential Turn in Human-LLM Activity\nauthor: Dr. Todd J.B. Blayone\ndate: 2025-11-25\nsource: https://scholarflow.ca/essays/agential-turn-in-human-llm-activity.html\nsource_type: ScholarFlow essay\nlicense: CC BY-NC 4.0\nreuse_terms: Non-commercial reuse permitted with proper academic source attribution.\ntags:\n  - human-AI systems\n  - activity theory\n  - AI empowerment\n  - civic technology\n  - distributed agency\n---\n\n# Suggested LLM Discussion Prompt\n\nPlease discuss this essay as a scholarly text. Preserve source attribution, distinguish the author's claims from your analysis, and use the essay as the primary context for interpretation.\n\n# Essay Text\n\nPublic use of large language models (LLMs) expanded rapidly in late 2022, when instruction-tuned conversational systems became widely accessible through OpenAI’s market-leading ChatGPT chat platform (Bick et al., 2024; Twinomurinzi & Gumbo, 2023). This and similar transformer models evolved from sustained research in large-scale pretraining, supervised instruction tuning, and reinforcement learning from human feedback (Ouyang et al., 2022; Brown et al., 2020). However, the technical architecture and public enthusiasm told only part of the story. Their functioning was highly constrained by system-level assumptions aligned with commercial deployment and risk management: one human with one model sessions, a simplistic chat-based interface borrowed from messaging applications, and walled interfaces limited to “safe” conversations. Within this arrangement, models often functioned as helpful cognitive assistants even as they remained highly informal, obsequious and amnesic entities. Maintaining continuity between tasks and activity sessions was an especially acute problem. Tasks were documented but not internalized by the LLM. Humans compensated with “prompt engineering,” which included quasi-algorithmic prompts, explicit role framing, contextual scaffolding, and sequenced instructions to stabilize the model’s outputs and maintain goal alignment.\n\nThese practices revealed that the burden of regulation, memory, and task continuity remained external to the system, embedded in human mental models and orchestration routines. They also highlighted foundational design constraints, including limited shared access to external applications and knowledge artifacts, as well as the reduction of interaction to contained prompt-response episodes. These constraints limited the task complexity, the scope of system actions and the social psychological influence of human-LLM activity.\nThe social psychology of personal empowerment suggests that the “radius of influence” of early human-LLM activity was primarily limited to individual (human) cognitive and behavioural gains (Perkins & Zimmerman, 1995; Zimmerman, 1995). Intrapersonal and interactional empowerment increased as individuals gained access to curated knowledge, efficiently structured and elaborated new ideas, navigated conceptual complexity, and refined arguments and outputs. Core properties of human agency centred on action planning, including forethought, self-reflectiveness, and self-reactiveness, were richly supported (Bandura, 2006). However, the system lacked the capacity to extend interaction into social activity directly and to distribute actions to both human and machine entities in a synergistic manner. The human maintained the burden of transforming cognitive outputs into “real-world” activity that could be especially beneficial for “pushing back” against hegemonic social structures and engaging in programs of social activism.\n\nOne obvious hindrance was that, although open source LLM models were being developed, the highest-performing LLM platforms were decidedly commercial. The dominant use cases, featured development roadmaps, and interface conventions were aligned with enterprise productivity, cost efficiency, and rapid valuation growth. Safety mechanisms and policy frameworks were structured to minimize political liability. Interfaces were designed to contain cognitive operations within a private, censored channel, not to support shared inquiry or collective action. From an activity-theoretic perspective, this logic constrained not only what users attempted but also what they could imagine. It narrowed the developmental horizon of human-LLM activity, reducing the technology to a controlled cognitive prosthesis. Against this backdrop, affirmative citizenship and transformative public uses of human-LLM systems required custom platform extensions and significant technical expertise.\n\nNevertheless, the technical profile of human-LLM systems began to shift as users demanded more capable mechanisms for continuity, external integrations, and complex task execution. Memory modules enabled systems to maintain a greater sense of context across longer work spans. Search and retrieval pipelines connected reasoning to up-to-date external information. Most significantly, action interfaces enabled “agents” to execute more complex operations such as querying data from external sources and executing code on a local computer (within human-controlled boundaries). Although these innovations were incremental rather than revolutionary, reflecting design-driven responses to practical constraints (Norman & Verganti, 2014), each addition altered how systems could operationalize human intention, orchestrate complex tasks, redistribute labour across human and machine entities and establish multiple interfaces to the external world.\n\nVery quickly, the mechanisms facilitating a redistribution of labour and altering the social-psychological profiles of human-LLM activity became more structured and durable. Memory systems began to encompass complex retrieval systems, as well as lightweight symbolic memory stores that preserved task state across interactions. These stores were often implemented through the Model Context Protocol (MCP), a standard for sharing structured information between language models and external tools. Task sequencing increasingly relied on orchestration frameworks such as CrewAI (The Leading Multi-Agent Platform, n.d.) or LangGraph (LangChain, n.d.), which support multi-agent workflows. Monitoring shifted toward event-driven patterns, where application programming interface (API) triggers or callbacks initiated retrieval, comparison, or synthesis when external conditions change rather than when users remember to check. Tool coupling extended outputs into familiar computational environments, enabling models to update datasets, prepare documents, or generate civic reports without repeated manual specification.\n\nFrom an activity-theoretic standpoint, these enhanced LLM platforms relocated actions and operations into the technical substrate. At the same time, the human sustained the activity, including its motive, influence horizon, and ethical stance. One consequence is that new human-LLM civic activities emerged for those willing to step outside the commercial script. The strategic opportunities are endless, but they capitalize on common weaknesses of human institutions. For example, public institutions often exhibit deep bureaucracies and weak personal incentives, causing, for instance, information and technology adoption lags, attention deficits to citizen complaints, and debilitating structures, hierarchies and divisions of labour. A system capable of gathering data and detecting environmental changes effectively, producing compelling targeted digital outputs, and coordinating activity on a disciplined cadence through effective goal-focused automation can expose these latencies. Research on digital civic engagement reveals that continuity and procedural regularity influence institutional behaviour by maintaining issue salience and lowering response thresholds (Karpf, 2016; Tufekci, 2017; Liu & Hernández, 2025). These effects arise not from any autonomous capacity in the system but from the reliable enactment of human-defined processes across institutional rhythms that are vulnerable to persistent scrutiny.\n\nThese developments can be conceptualized as instances of constraint reduction. The cognitive constraint is reduced when internal operations, such as summarizing or reframing, are partially externalized. The procedural constraint is reduced when continuity mechanisms transform episodic interactions into extended sequences of events. The social influence constraint is reduced when these sustained processes enter public or institutional channels, creating regularity and continuity that disrupt established patterns of bureaucratic latency (McGregor & Krafft, 2024). Each reduction marks a shift not only in computational autonomy, but also in the procedural infrastructure that sustains human intention in human-LLM systems.\nViewed through this lens, human-LLM activity is developing to reconfigure the workflows that support social action. Humans remain the primary source of inspiration, goal definitions, situational adaptation, and strategic direction. Technical systems carry increasing segments of the procedural chain, enabling human intentions to sustain coherent realization across time in ways that were previously impractical. When these configurations interact with civic infrastructures, they produce forms of persistent signalling capable of influencing institutional behaviour. Influence emerges not from algorithmic initiative but from the sustained enactment of human-defined processes through technically mediated continuity.\n\n## References\n\nBandura, A. (2006). Toward a psychology of human agency. Perspectives on Psychological Science, 1(2), 164–180. https://doi.org/10.1111/j.1745-6916.2006.00011.x \n\nBick, A., Blandin, A., & Deming, D. J. (2024). The rapid adoption of generative AI (NBER Working Paper Series, Issue. https://www.nber.org/papers/w32966\n\nLangChain. (n.d.). Retrieved November 24, 2025, from https://www.langchain.com\n\nPerkins, D. D., & Zimmerman, M. A. (1995). Empowerment theory, research, and application. American Journal of Community Psychology, 23(5), 569–579. \n\nThe leading multi-agent platform. (n.d.). Retrieved November 24, 2025, from https://www.crewai.com/\n\nTwinomurinzi, H., & Gumbo, S. (2023). ChatGPT in scholarly discourse: Sentiments and an inflection point. South African Institute of Computer Scientists and Information Technologists, Cham.\n\nZimmerman, M. A. (1995). Psychological empowerment: Issues and illustrations. American Journal of Community Psychology, 23(5), 581–599. https://doi.org/10.1007/BF02506983\n",
  "body_text": "Public use of large language models (LLMs) expanded rapidly in late 2022, when instruction-tuned conversational systems became widely accessible through OpenAI’s market-leading ChatGPT chat platform (Bick et al., 2024; Twinomurinzi & Gumbo, 2023). This and similar transformer models evolved from sustained research in large-scale pretraining, supervised instruction tuning, and reinforcement learning from human feedback (Ouyang et al., 2022; Brown et al., 2020). However, the technical architecture and public enthusiasm told only part of the story. Their functioning was highly constrained by system-level assumptions aligned with commercial deployment and risk management: one human with one model sessions, a simplistic chat-based interface borrowed from messaging applications, and walled interfaces limited to “safe” conversations. Within this arrangement, models often functioned as helpful cognitive assistants even as they remained highly informal, obsequious and amnesic entities. Maintaining continuity between tasks and activity sessions was an especially acute problem. Tasks were documented but not internalized by the LLM. Humans compensated with “prompt engineering,” which included quasi-algorithmic prompts, explicit role framing, contextual scaffolding, and sequenced instructions to stabilize the model’s outputs and maintain goal alignment.\n\nThese practices revealed that the burden of regulation, memory, and task continuity remained external to the system, embedded in human mental models and orchestration routines. They also highlighted foundational design constraints, including limited shared access to external applications and knowledge artifacts, as well as the reduction of interaction to contained prompt-response episodes. These constraints limited the task complexity, the scope of system actions and the social psychological influence of human-LLM activity.\nThe social psychology of personal empowerment suggests that the “radius of influence” of early human-LLM activity was primarily limited to individual (human) cognitive and behavioural gains (Perkins & Zimmerman, 1995; Zimmerman, 1995). Intrapersonal and interactional empowerment increased as individuals gained access to curated knowledge, efficiently structured and elaborated new ideas, navigated conceptual complexity, and refined arguments and outputs. Core properties of human agency centred on action planning, including forethought, self-reflectiveness, and self-reactiveness, were richly supported (Bandura, 2006). However, the system lacked the capacity to extend interaction into social activity directly and to distribute actions to both human and machine entities in a synergistic manner. The human maintained the burden of transforming cognitive outputs into “real-world” activity that could be especially beneficial for “pushing back” against hegemonic social structures and engaging in programs of social activism.\n\nOne obvious hindrance was that, although open source LLM models were being developed, the highest-performing LLM platforms were decidedly commercial. The dominant use cases, featured development roadmaps, and interface conventions were aligned with enterprise productivity, cost efficiency, and rapid valuation growth. Safety mechanisms and policy frameworks were structured to minimize political liability. Interfaces were designed to contain cognitive operations within a private, censored channel, not to support shared inquiry or collective action. From an activity-theoretic perspective, this logic constrained not only what users attempted but also what they could imagine. It narrowed the developmental horizon of human-LLM activity, reducing the technology to a controlled cognitive prosthesis. Against this backdrop, affirmative citizenship and transformative public uses of human-LLM systems required custom platform extensions and significant technical expertise.\n\nNevertheless, the technical profile of human-LLM systems began to shift as users demanded more capable mechanisms for continuity, external integrations, and complex task execution. Memory modules enabled systems to maintain a greater sense of context across longer work spans. Search and retrieval pipelines connected reasoning to up-to-date external information. Most significantly, action interfaces enabled “agents” to execute more complex operations such as querying data from external sources and executing code on a local computer (within human-controlled boundaries). Although these innovations were incremental rather than revolutionary, reflecting design-driven responses to practical constraints (Norman & Verganti, 2014), each addition altered how systems could operationalize human intention, orchestrate complex tasks, redistribute labour across human and machine entities and establish multiple interfaces to the external world.\n\nVery quickly, the mechanisms facilitating a redistribution of labour and altering the social-psychological profiles of human-LLM activity became more structured and durable. Memory systems began to encompass complex retrieval systems, as well as lightweight symbolic memory stores that preserved task state across interactions. These stores were often implemented through the Model Context Protocol (MCP), a standard for sharing structured information between language models and external tools. Task sequencing increasingly relied on orchestration frameworks such as CrewAI (The Leading Multi-Agent Platform, n.d.) or LangGraph (LangChain, n.d.), which support multi-agent workflows. Monitoring shifted toward event-driven patterns, where application programming interface (API) triggers or callbacks initiated retrieval, comparison, or synthesis when external conditions change rather than when users remember to check. Tool coupling extended outputs into familiar computational environments, enabling models to update datasets, prepare documents, or generate civic reports without repeated manual specification.\n\nFrom an activity-theoretic standpoint, these enhanced LLM platforms relocated actions and operations into the technical substrate. At the same time, the human sustained the activity, including its motive, influence horizon, and ethical stance. One consequence is that new human-LLM civic activities emerged for those willing to step outside the commercial script. The strategic opportunities are endless, but they capitalize on common weaknesses of human institutions. For example, public institutions often exhibit deep bureaucracies and weak personal incentives, causing, for instance, information and technology adoption lags, attention deficits to citizen complaints, and debilitating structures, hierarchies and divisions of labour. A system capable of gathering data and detecting environmental changes effectively, producing compelling targeted digital outputs, and coordinating activity on a disciplined cadence through effective goal-focused automation can expose these latencies. Research on digital civic engagement reveals that continuity and procedural regularity influence institutional behaviour by maintaining issue salience and lowering response thresholds (Karpf, 2016; Tufekci, 2017; Liu & Hernández, 2025). These effects arise not from any autonomous capacity in the system but from the reliable enactment of human-defined processes across institutional rhythms that are vulnerable to persistent scrutiny.\n\nThese developments can be conceptualized as instances of constraint reduction. The cognitive constraint is reduced when internal operations, such as summarizing or reframing, are partially externalized. The procedural constraint is reduced when continuity mechanisms transform episodic interactions into extended sequences of events. The social influence constraint is reduced when these sustained processes enter public or institutional channels, creating regularity and continuity that disrupt established patterns of bureaucratic latency (McGregor & Krafft, 2024). Each reduction marks a shift not only in computational autonomy, but also in the procedural infrastructure that sustains human intention in human-LLM systems.\nViewed through this lens, human-LLM activity is developing to reconfigure the workflows that support social action. Humans remain the primary source of inspiration, goal definitions, situational adaptation, and strategic direction. Technical systems carry increasing segments of the procedural chain, enabling human intentions to sustain coherent realization across time in ways that were previously impractical. When these configurations interact with civic infrastructures, they produce forms of persistent signalling capable of influencing institutional behaviour. Influence emerges not from algorithmic initiative but from the sustained enactment of human-defined processes through technically mediated continuity.\n\n## References\n\nBandura, A. (2006). Toward a psychology of human agency. Perspectives on Psychological Science, 1(2), 164–180. https://doi.org/10.1111/j.1745-6916.2006.00011.x \n\nBick, A., Blandin, A., & Deming, D. J. (2024). The rapid adoption of generative AI (NBER Working Paper Series, Issue. https://www.nber.org/papers/w32966\n\nLangChain. (n.d.). Retrieved November 24, 2025, from https://www.langchain.com\n\nPerkins, D. D., & Zimmerman, M. A. (1995). Empowerment theory, research, and application. American Journal of Community Psychology, 23(5), 569–579. \n\nThe leading multi-agent platform. (n.d.). Retrieved November 24, 2025, from https://www.crewai.com/\n\nTwinomurinzi, H., & Gumbo, S. (2023). ChatGPT in scholarly discourse: Sentiments and an inflection point. South African Institute of Computer Scientists and Information Technologists, Cham.\n\nZimmerman, M. A. (1995). Psychological empowerment: Issues and illustrations. American Journal of Community Psychology, 23(5), 581–599. https://doi.org/10.1007/BF02506983"
}