thinkingmachine.xyzPHILOSOPHY + PSYCHOLOGY
How can people and AI agents cooperate to understand the world, improve engineering systems and pursue goals we can stand behind?
OUR STARTING POINT
We believe people and AI agents should work together towards common goals, combining different capabilities and helping one another understand the work. A useful collaboration leaves the team better informed, with a clearer next step and a shared account of what happened.
Here, an agent is a contributor able to pursue a task within an agreed scope. People bring purpose, experience, values and responsibility; software agents can investigate, organise, propose and check. Calling both contributors does not imply that they have the same understanding, consciousness or moral status.
This is ThinkingMachine’s developing position. The external readings below inform the discussion; their authors do not endorse our services or validate our implementation.
THE PSYCHOLOGY OF COLLABORATION
Risko and Gilbert [5] review how external tools reduce cognitive demands. Our assessments of when to offload can also be mistaken. In an engineering workflow, delegate repetitive gathering while keeping the evidence available for interpretation and challenge.
Lee and See [6] explain the role of trust in reliance on automation. A confident answer is not a performance measure. Show what was checked, where the system has been evaluated and what remains uncertain, so people can judge when to rely on it.
Bainbridge [7] describes how automation can leave people with demanding exceptional situations. Give reviewers enough context, time and skill to intervene. A person asked to approve an opaque result has little opportunity to exercise meaningful judgement.
A 2025 survey of 319 knowledge workers [8] associated greater confidence in generative AI with less reported critical-thinking effort, while identifying verification, integration and task stewardship as important activities. It was a self-report survey: it does not establish that AI causes lasting loss of thinking skills. Our design response is to preserve opportunities to question, compare and learn.
WATCH, THEN DISCUSS
In Why AI Agents are either the best or worst thing we’ve ever built (1 May 2026), Hannah Fry and Brendan Maginnis experiment with an AI agent given real-world tasks and a bank card. Fry’s published description reports a novelty-mug shop, an unsolicited email to a journalist and leaked passwords.
The film credits Nicklas Berild Lundblad, whose writing on abundant agency considers how increased capacity for goal-directed action could change institutions and accountability. His essay argues for complementary human and synthetic agency while protecting human autonomy.
The film is an illustrative experiment, not a controlled study of all agents. The essay is philosophical commentary. Our interpretation below is separate from both sources.
Watch Hannah Fry’s filmOUR TAKEAWAY
Common goals need more than a short instruction. They need constraints, a way to resolve disagreement and a record of who authorised what.
An agent might achieve a narrow target while creating costs or consequences the team never intended. Define success alongside budgets, information boundaries and the interests of other people.
Useful cooperation also includes restraint: report uncertainty, ask for a decision when needed and stop when the approved scope ends.
FROM PHILOSOPHY TO ENGINEERING
Record the desired outcome, affected people, acceptance evidence and limits. Keep a named human owner responsible for decisions and resolving conflicts.
Use scoped MCP tools to access approved information. A shared interface helps contributors participate; identities and service-side permissions define what each may do.
Record sources, uncertainty and alternative explanations. Another agent can challenge a proposal, but may share its errors. Review against underlying evidence.
Compare the result with the agreed goal. Record corrections and lessons, then review prompt and workflow changes before release. Leave the next contributor a usable handoff.
Our OpenProject example puts this into a managed work item: a person defines the outcome, agents contribute attributed findings and proposals, and the owner records decisions. Prompt governance keeps versions and evaluations connected to that record. These are our proposed engineering applications, not outcomes demonstrated by the cited papers.
REFERENCES + FURTHER READING
Start with Hannah Fry’s film and Clark’s 2025 article, then explore the psychology papers. Earlier studies provide foundations; they did not test today’s MCP-connected LLM agents.
Andy Clark & David Chalmers · 1998 · Philosophy · Author-hosted text. The foundational argument about cognition extending through external resources.
Andy Clark · 2025 · Philosophical commentary · Open access. Hybrid thinking, judgement and the care needed when incorporating AI into cognitive activity.
Edwin Hutchins · 1995 · Cognitive science · Publisher access may be required. An analysis of remembering across people and cockpit equipment.
Luciano Floridi · 2023 · Philosophical editorial · Publisher access required. A distinction between artificial agency and intelligence, applied to generative models.
Evan Risko & Sam Gilbert · 2016 · Psychology review · Free author-hosted PDF. How and why people use external resources to reduce cognitive demands.
John Lee & Katrina See · 2004 · Human factors review · Publisher access required. How trust, context and system characteristics influence reliance.
Lisanne Bainbridge · 1983 · Engineering human factors · Publisher access may be required. Why automation can leave difficult problems for human operators.
Hao-Ping Lee and colleagues · 2025 · Self-report survey · Research page with paper. Confidence, cognitive effort and changing critical-thinking practices in knowledge work.
Saleema Amershi and colleagues · 2019 · Evaluated design guidelines · Free PDF. Practical guidance for capabilities, errors, correction and changing behaviour.
Fabrizio Dell’Acqua and colleagues · Research first released 2023 · Field experiment · Institutional overview linking the paper. AI usefulness varies across consulting tasks; engineering outcomes need their own evaluation.
Hannah Fry · 2026 · Documentary experiment · YouTube. Featuring Brendan Maginnis and Nicklas Berild Lundblad; a starting point for discussing delegation and consequences.
Nicklas Berild Lundblad · 2025 · Author’s essay. See the “Abundant Agency” section for complementarity, oversight and human autonomy.
Reading list reviewed 10 October 2026. Summaries are paraphrases; access conditions may change. Following a link opens the external publisher’s site. No external video or tracking embed is loaded here.
LET’S FIND YOUR FIRST WORKFLOW