Working with Agentic `Teammates': When a New Organizational Actor Collides with the Human Ecosystem of Work
Rida Qadri, Remi Denton, Michael Madaio, Mahima Pushkarna, Leslie Lai, Sherry Moore, Michelle Chen Huebscher, Andrew Butcher, et al.
cs.HC · cs.AI
Quick verdict
In-situ study (n=17) of a proactive AI teammate in a real company: norms, relational boundaries and agency in flux; rich but small, single-firm, self-reported.
Summary
This paper reports an in-situ qualitative study of 'Team Agent,' a persistent, proactive LLM-based agent with a dedicated service account, deployed across 20+ teams in a large technology company for five months (over 41,000 conversational turns, 11,000+ agent responses). Framed as a technology probe rather than a product evaluation, the study draws on semi-structured interviews with 17 participants across 11 teams — mostly software engineers, with usage spanning one week to five months — analyzed through reflexive thematic analysis. The authors find three collisions: (1) the agent failed at tacit collaborative norms, citing stale documents as authoritative, misreading humor and brainstorming, and tagging a VP in 16 bugs, thereby creating remedial work; (2) users held fragmented, polarized mental models of the agent's ontological status and relational boundaries, especially around personas, praise, and interpersonal feedback; and (3) acceptance of proactivity hinged on perceived agency — forced adoption and surveillance feelings chilled team chat spaces, and participants expected autonomy to be progressively earned. The authors argue for treating agents as a new non-human organizational actor and recommend earned autonomy, team-negotiated codes of conduct, agent-aware infrastructure, and consent-based deployment. Key limitations: a single tech-forward company, purely self-reported data, a snapshot rather than longitudinal design, and an insider research context where the employer also built the agent; sample selection and response rates are not reported.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Proactive team-embedded agents deployed into live organizations collide with tacit collaborative norms, generating noise and remedial work (§4.1).
- Users lack a stable category for such agents and negotiate ontological status and relational boundaries in fragmented, polarized ways (§4.2).
- Acceptance of agent autonomy is mediated by perceived human agency; proactivity is expected to be an earned, progressive privilege rather than an out-of-the-box capability (§4.3).
- Agents should be conceptualized as a novel non-human organizational actor, demanding new governance structures, progressive-autonomy design, and agent-aware collaborative infrastructure (§5).
Strengths
Rare real-world deployment evidence
Unlike the lab, Wizard-of-Oz, and game-bot studies the paper critiques (§2.1), this is a five-month deployment across 20+ teams with 41,000+ conversational turns and 11,000+ agent responses (§3.2.3) — ecological validity that simulated setups cannot offer.
Concrete, well-sourced qualitative findings
Each theme is anchored in vivid episodes: stale documents treated as 'law of the land' (P4, §4.1.1), a VP tagged in 16 bugs (§4.1.1), the 'bystander effect' where agent reactions suppress human help (P4, §4.3.3), and tone-policing DMs that chill chat (P2, §4.3.3).
Credible probe framing and governance separation
Actionable design and governance agenda
Weaknesses
Opaque sampling from an unknown population
Self-report despite owning behavioral logs
Insider research and conflict-of-interest exposure
No analytic audit trail
Version hygiene issues
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This paper reports one of the first in-situ qualitative studies of a genuinely deployed, proactive, multi-user AI agent. 'Team Agent' — an LLM-based system on a dedicated service account with multi-surface persistence across chat, documents, bug trackers, and meetings — ran for five months across 20+ teams inside a single large technology company, accumulating 41,000+ conversational turns and 11,000+ agent responses (§3.2.3). The authors explicitly position the system as a technology probe (§3.3), not a product evaluation: the goal is to expose what happens when a maximally proactive, socially personified agent enters a human ecosystem of work. Three claims structure the paper (Table 3): (1) capable agents still violate tacit collaborative norms, converting helpfulness into noise and remedial labor (§4.1); (2) users negotiate the agent's ontological status and relational boundaries in fragmented, polarized ways (§4.2); (3) acceptance of agent autonomy is gated by perceived human agency, with autonomy expected to be 'earned' progressively rather than asserted (§4.3). The discussion (§5.1–5.2) elevates these into a conceptual proposal — agents as a novel non-human organizational actor — plus design recommendations: progressive autonomy, team-negotiated codes of conduct, agent-aware infrastructure, and consent-based deployment.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is 'Team Agent' in this paper?
A persistent, proactive LLM-based AI agent with its own service account that was embedded in the chats, documents, and bug trackers of 20+ teams at a large technology company for five months. It could file bugs, schedule meetings, summarize threads, and even coach interpersonal communication — acting on its own initiative rather than only responding to commands.
What did the study find about AI teammates at work?
Interviews with 17 users across 11 teams found three friction zones: the agent violated unwritten collaboration norms (citing outdated docs, tagging executives, misreading jokes); users held conflicting views on what the agent socially 'is' and how to interact with it; and people accepted its autonomy only when they felt in control, expecting trust to be earned gradually. Forced adoption and surveillance feelings led some teams to avoid shared chat spaces.
Which company was studied — is it Google?
The paper anonymizes the organization, but all authors are Google Research or Google DeepMind employees, so readers typically infer the deployment happened at Google. The paper itself never names the company.
What do the authors recommend for deploying AI agents in teams?
Progressive ('earned') autonomy instead of day-one proactivity, team-negotiated codes of conduct defining appropriate agent behavior, consent-based team-level deployment rather than top-down mandates, and retooled collaboration software with agent-specific permissions, audit trails, and rollback mechanisms.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Tables 1–3 were flattened by text extraction; their structure was reconstructed from surrounding prose.
- Provided metadata lists 8 authors while the extracted byline lists 16; the review treats the extracted text as authoritative.
- arXiv preprint with placeholder template dates (2007/2009) in the backmatter; venue peer-review status is unknown.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.