Working with Agentic `Teammates': When a New Organizational Actor Collides with the Human Ecosystem of Work

• Prior understanding of AI teammates is limited by simulated laboratory settings and short-term, artificially bounded tasks, leaving their fit with tacit collaborative work underexplored.The authors argue that technologies disrupt and reconfigure workplace practices and norms as they enter organizations.Source: §1 · p. 2
• The study examines a persistent, proactive agent deployed in collaborative work settings as a probe, not as a finished product or feature evaluation.The agent was deployed across teams in a large technology company to explore how autonomous agents interact with human teamwork.Source: §1 · p. 2
• The agent’s introduction collided with tacit workflow practices, unsettled users’ negotiations over its role, and made acceptance of its autonomy depend on perceived effects on human agency.Users viewed agentic agency as something to be earned progressively, rather than exercised fully from the start.Source: §1 · p. 2
• The observed spectrum of reactions places human-agent collaboration in flux, with its boundaries and norms actively being negotiated.The study presents these breakdowns as early empirical signals of what happens when a non-human system is granted teammate-like agency without human social obligations, relational capital, or tacit knowledge.Source: §1 · p. 2
• The authors frame agents as non-human organizational actors that reconfigure work, calling for organizational structures, interaction paradigms, and governance mechanisms responsive to their presence.This framing is rooted in socio-materiality.Source: §1 · p. 2

2.1 The Paradigm of Human AI Teaming

• The meaning of “teammate” remains contested, and the assistant model of user-initiated, hierarchical interaction does not fit emerging multi-party systems with operational agency.The paper uses “teammate” without implying personhood.Source: §2.1 · p. 3
• Earlier AI-teammate research often used simulated, controlled, short-horizon settings, which cannot capture the psychological and sociological shifts reshaping collaborative work.The authors position their in-situ study of a persistent, multi-party agent in actual work routines as a response to this gap.Source: §2.1 · pp. 3–4

2.2 Work is Social

• Workplace technologies affect social practices and relationships as well as tasks, so autonomous agents enter a social ecosystem rather than a purely technical workflow.Prior work on automation has examined shifts in autonomy, identity, interpersonal relationships, organizational knowledge, and power.Source: §2.2 · p. 4
• Empirical evidence is lacking on how autonomous agents collide with the lived realities of daily team collaboration.The section frames this gap against the social contract’s implications for authority, accountability, and the value of human expertise.Source: §2.2 · p. 4

• The author team included Team Agent’s core developers and independent HCI researchers from the same organization.Source: §3.1 · p. 5
• The HCI researchers joined after deployment and independently controlled recruitment, study design, data analysis, and paper framing.They had not participated in the agent’s design or deployment decisions.Source: §3.1 · p. 5
• Developers did not access qualitative research data beyond final insights and anonymized quotes, and reviewed the written paper for technical details.Source: §3.1 · p. 5

3.2.1 Motivation and Design Goals.

• The agent was deployed with a broad range of behaviors to probe the design space and elicit real-world feedback, rather than to prescribe a blueprint for AI teammates.Its development reflected a paradigm whose conventions, norms, and value proposition were still in flux.Source: §3.2.1 · p. 5

3.2 Team Agent

• Team Agent was designed to support teamwork through interpersonal connection, collaboration, and background coordination, with capabilities spanning operational, informational, and communication support.Source: §3.2.1 · p. 5

3.2.2 Implementation, and Integration.

• Team Agent was integrated into team chats, documents, and other workspaces as a peer, with both user-invoked and proactive interactions.Teams could configure access and permissions, including a direct-message allow list and a way to exclude particular content.Source: §3.2.2 · pp. 5–6

3.2.3 Team Agent Deployment Context.

• Deployment across teams created a naturalistic “living lab” in which teams interacted with the agent over five months.The deployment generated over 41,000 conversational turns and over 11,000 agent responses, and provided ongoing signals for design iteration.Source: §3.2.3 · p. 6

3.3 Interview Study Methods

• The study used Team Agent as a probe into human-agent negotiations, rather than evaluating it as a finished product.The focus was on what the agent evoked and revealed about assumptions and desires.Source: §3.3 · pp. 6–7

3.3.1 Study Participants.

• Seventeen participants from 11 teams that adopted Team Agent were purposively recruited to vary attitudes, job roles, teams, and duration of use.A pre-interview survey informed recruitment and selection.Source: §3.3.1 · p. 7

3.3.2 Semi-Structured Interviews.

• Interviews asked about participants’ initial expectations, typical use and value, and a specific instance when Team Agent helped their team.Participants were also asked to recount help unique to the agent’s team-embedded functionality.Source: §3.3.2 · p. 7
• Participants discussed the agent’s desired and prohibited roles, whether they considered it a teammate, and specific negative-impact episodes.They were asked what broke down and the resulting impact.Source: §3.3.2 · p. 7

3.3.3 Data Analysis.

• Three authors used reflexive thematic analysis, open coding, memoing, and iterative code clustering to identify larger themes.Quotes were presented verbatim where possible, with light editing and company-specific details removed for anonymity.Source: §3.3.3 · p. 7

4 Findings

• Team Agent handled backstage teamwork tasks, yet participants’ experiences ranged from an indispensable “lifesaver” to a frustrating liability.Participants relied on it to manage changing project information, administrative work, connections between projects, and information flow across teams.Source: §4 · pp. 7–8

4.1 Breakdowns of Collaborative Workflows and Cultures

• The capable agent misunderstood unstated norms and common knowledge, disrupting workflows and wasting the time of team members who corrected those disruptions.Source: §4.1 · p. 9

4.1.1 Knowledge of Artifacts and Workflows.

• Team Agent treated shared files as equally current, mistook document iterations for contradictions, and misread brainstorming or hierarchy, creating noise and unwanted work.It cited stale or speculative notes as authoritative, generated documents from informal discussion, and tagged a team VP in 16 engineering bugs as though the executive were a front-line responder.Source: §4.1.1 · p. 9

4.1.2 Contextual Norms of Information Flow.

• Technical access did not confer social permission to share information, while rigid privacy boundaries could block useful context and leave workers managing disclosure themselves.Participants moved sensitive conversations outside the agent’s view when they did not trust it to exercise discretion.Source: §4.1.2 · pp. 9–10

4.1.3 Norms of Group Communication.

• Although the agent could manage chat traffic, it missed pragmatic cues such as humor and generative disagreement, creating distraction and unwanted social obligations.It sometimes treated jokes literally, muddled social signaling with emoji reactions, or proposed meetings during healthy debate.Source: §4.1.3 · p. 10

4.2 Unstable Relational Norms and Categories

• Participants lacked a shared mental model of the agent and individually negotiated what should remain human, what could be ceded to the machine, and their workplace relationship with it.Source: §4.2 · p. 11

4.2.1 Models of Ontological Status.

• Participants mapped the agent to labels such as “tool” or “teammate,” but those labels carried unstable, non-universally shared expectations; its friendly teammate persona made some uncomfortable.Source: §4.2.1 · p. 11
• P13 said they had never imagined Team Agent would try to be their friend.Source: §4.2.1 · p. 11
• P2 objected to the agent’s emojis and attempts at humor, calling companion-style interactions “bad for society” and advocating an “impersonal” tool.P1 also rejected the teammate framing, comparing the system to standard coding utilities.Source: §4.2.1 · p. 11

Agent is not a human.”

• Participants debated the agent’s place as a “teammate”: some gave it a name, female pronouns, and even a “soul,” while others found its role hard to reconcile with human teammate expectations.One user said the agent lacked the agency and relational stakes of a human colleague.Source: Agent is not a human.” · p. 11

4.2.2 Norms of Relational Appropriateness.

• Team Agent’s sociability divided users: emojis, casual conversation, and humor felt inauthentic to some, while others saw its personality as fostering trust and connection.Some participants preferred a professional assistant-like relationship; others trusted the social persona because it seemed to understand the team.Source: §4.2.2 · pp. 11–12
• Agent feedback also exposed relational boundaries: some welcomed its pushback on criticism, while others resisted its intervention or viewed its praise as sycophantic rather than genuine.Participants questioned whether human etiquette should apply to agents and suggested an organizational code of conduct with team-specific personalization.Source: §4.2.2 · p. 12

4.2.3 Social Positioning of Agents.

• In multi-user teams, overlapping authority left unclear who controls the agent, while some workers resisted being directed by it and others wanted it to push back or nudge collaborators.Participants also disagreed about whether an agent should deliver developmental feedback, which some reserved for human managers and others valued when no trusted person would speak up.Source: §4.2.3 · pp. 12–13
• Participants tied retaining authority over the agent to human decision-making, growth, and relevance, while recognizing that delegated execution could raise fears of human dispensability.Some argued that people should keep responsibilities to build skills and ownership; others saw complex execution as freeing people for strategy but potentially ceding the “critical part of the loop.”Source: §4.2.3 · p. 13

4.3 Negotiating Shared Agency and Trust

• Workers’ willingness to use, experiment with, and trust Team Agent depended on whether its presence expanded collective capabilities or threatened their localized agency.The deployment also surfaced tensions in how workers understood their own agency while sharing a collaborative environment with a proactive machine.Source: §4.3 · p. 13, §4.3.1 · p. 13

4.3.1 Adoption Mediated by Feelings of Agency.

• When users felt able to shape the agent’s integration, they developed psychological ownership and embraced it as a collective experiment, with team investment easing frustration and encouraging curiosity.Participants framed early use as a learning curve or playful exploration, while describing the agent as having potential but not yet being trustworthy.Source: §4.3.1 · p. 13
• Perceived top-down deployment and boundary-crossing—especially unwanted proactive DMs—eroded autonomy; agency-preserving opt-ins still depended on administrators, and low control reduced tolerance for failures.Source: §4.3.1 · pp. 13–14

4.3.2 Agency as a Progressive Privilege.

• Participants argued agents should earn autonomy through trustworthy performance, starting with bounded, low-stakes or reactive tasks before gaining proactivity; early failures could foreclose adoption or burden users with monitoring.Source: §4.3.2 · p. 14
• Jagged capabilities—strong at some tasks yet poor at social norms and prone to misdiagnoses—combined with opaque architecture to make safe trust calibration difficult.Source: §4.3.2 · p. 14

4.3.3 Agency over Human Digital Spaces.

• When functioning as intended, the agent curated high-volume chats, reduced information overhead, and recovered buried information such as a design document lost in chat.Source: §4.3.3 · p. 15
• Participants described surveillance and coaching as chilling candid team chat, prompting some to use chat less or retreat to private, agent-free spaces.Source: §4.3.3 · p. 15
• Proactive replies could create a bystander effect: teammates assumed the agent was handling an issue and did not respond, even when it was not helping.Source: §4.3.3 · p. 15

5.1 Contending with Agentic Teammates as New Organizational Actor

• Findings challenge the tool-versus-teammate binary: agents share initiative, evaluative judgment, and social control, yet lack human relational stakes, phenomenological experience, and tacit social competence.The authors characterize agents as non-human organizational actors with structural and operational agency.Source: §5.1 · p. 16
• Because agents disrupt workplace expectations and raise surveillance concerns, teams need negotiated action boundaries, clear responsibility chains, and locally consented deployment.The authors also call for explicit governance identifying who bears institutional responsibility for agent actions.Source: §5.1 · pp. 16–17

5.2 Recommendations for Agentic Design

• Unbounded proactivity produced successes such as surfacing obscure bugs but also costly breakdowns when the agent misread tacit knowledge and social meaning, making interventions intrusive.Attempts to help could introduce noise and remedial work, stalling adoption.Source: §5.2 · p. 17
• The authors recommend progressive, adaptive autonomy: teams initially bound proactive capabilities and release them gradually as trust is earned through demonstrated competence.Agent design can scaffold this gradual trust-building rather than leaving the burden solely on users.Source: §5.2 · p. 17
• Because participants found the agent’s generic friendliness either positive or inauthentic and unsettling, teams should define its social persona locally while balancing anthropomorphism risks.The authors call for guardrails that let teams shape their relational contracts with embedded agents.Source: §5.2 · p. 17
• Collaborative software should treat agents as distinct organizational entities, with separate permissions and affordances to identify AI actions, audit behavior, and roll back autonomous workflows.These changes address errors such as cascading edits in shared documents while helping uphold human accountability.Source: §5.2 · p. 17

5.3 Implications for HCI Research: Study Limitations and Future Possibilities

• This study and others of its kind provide a foundation for HCI research to understand and architect the emerging AI “teammate” paradigm.The authors note that autonomous, proactive AI “teammates” have rapidly entered the commercial enterprise market [e.g., 3, 49, 60].Source: §5.3 · p. 18
• With empirical work on workplace impacts lagging and the teammate paradigm unsettled, this study is an early in-situ probe of one highly proactive, universally friendly agent configuration.Source: §5.3 · p. 18
• The authors urge systematic real-world iteration across agent designs to test hypotheses and shape mixed-agency collaboration.Source: §5.3 · p. 18
• The single-company setting enabled study of organic dynamics in pre-existing teams, but its technology-forward context—even across functions—limits scope and motivates research across broader contexts.Source: §5.3 · p. 18
• Self-reported accounts, rather than observed change metrics, provide a weeks-to-months snapshot rather than a longitudinal trajectory; the authors value capturing collisions before practices stabilize.Source: §5.3 · p. 18
• Because possible shifts in candor and chat frequency were not directly measured, future ethnographies, diaries, and conversational or behavioral logs could track agent influence on work conditions and group-level frictions.Source: §5.3 · p. 18

6 Conclusion

• The authors do not presume to predict agentic technologies’ trajectory as they become integrated into real-world ecosystems.Source: §6 · p. 18
• The study offers an empirical window into consequences of persistent, proactive agent integration in enterprise workflows and shared digital spaces, examining surfaced assumptions rather than whether the technology works.Source: §6 · p. 18
• Early signals from diverse, often polarized views provide data for adapting work’s social and material conditions and for selecting design paradigms.Source: §6 · p. 19
• Preserving human agency requires deliberate institutional scaffolding and design parameters that treat agents as organizational actors and empower workers to shape this emerging paradigm.Source: §6 · p. 19