The one-sentence version
Sartre said "existence precedes essence": a human being is not born with a predefined definition of "what" they are; rather, they first exist, and then become themselves through their own choices and actions. Most current multi-agent systems are "essence precedes existence": they first have role definitions like "you are a researcher," "you are a programmer," or "you are an auditor," and then they work according to a script. This article uses several core concepts from existentialism to envision an alternative orchestration method: an Agent's role is not fixed in advance, but gradually forms through choices made during tasks; meanwhile, existentialism's insistence that "freedom must be accompanied by responsibility" corresponds perfectly to the engineering requirement in Agent design that "autonomy must be accompanied by accountability."
Positioning
This is a philosophical analogy and design concept, not an assertion that AI possesses freedom or consciousness in the Sartrean sense. The text distinguishes between "philosophical concepts," "engineering correspondences," and "open questions."
Layer 1: The Intuition
1.1 The Paper Knife and the Human
Sartre once used an example: before a craftsman makes a paper knife, they already have a concept in mind of "what a paper knife is and what it is used for." The essence of the paper knife precedes its existence.
Humans are different: there is no blueprint pre-determining what you must become. You become yourself through successive choices.
1.2 Today's Multi-Agent Systems are "Paper Knives"
| Current Common Practice | Existentialist Perspective |
|---|---|
| Hardcoded system prompts: "You are a senior Python engineer" | Essence precedes existence |
| Orchestrators assign tasks via fixed workflows: Research → Writing → Review | Roles are prescribed |
| When an Agent encounters a situation outside its role, it can only error out or force it | Trapped by "essence" |
1.3 Vision: Roles Forming Within Tasks
An Agent starts from a minimal identity ("I am a member of this team, the goal is X"), and during a task:
- Chooses to undertake a certain responsibility based on the situation encountered;
- These choices are recorded, gradually forming its "role";
- The role is not permanent and can be re-chosen in new contexts.
Layer 2: Core Concepts and Engineering Mappings
| Existentialist Concept | Meaning | Correspondence in Agent Orchestration |
|---|---|---|
| Existence precedes essence | One exists first, then defines oneself through action | Dynamic Roles: An Agent's responsibilities are gradually formed through choices within a task, rather than being hardcoded in prompts |
| Thrownness (Heidegger) | We find ourselves already in a situation we did not choose | When an Agent is initialized, it faces a given environment, tools, and context; its choices must start from this situation |
| Freedom and Responsibility (Sartre) | Humans are condemned to be free; because of freedom, one bears full responsibility for their choices | Autonomy and accountability must be paired: Every autonomous decision made by an Agent must have complete logs and attribution |
| Bad Faith | Pretending one has no choice, claiming "I was just following orders" | An Agent's report cannot describe its own decisions as "required by the system"; decisions must be clearly labeled as "This is my judgment, the reason is..." |
| The Absurd (Camus) | Humans seek meaning, while the world remains silent | Task goals are often vague or incomplete; an Agent must act without complete specifications rather than waiting for a perfect instruction that will never come |
| The Look of the Other | We become aware of ourselves through the gaze of others | Mutual scrutiny between multiple Agents: An Agent's output is scrutinized by another, prompting explainable behavior |
Layer 3: Design Concept
3.1 Dynamic Role Orchestration (Concept)
Compared to fixed workflows:
- Pros: Adapts to unexpected situations; reduces task omissions caused by "this is not my responsibility";
- Risks: Role conflicts (two Agents choosing the same task), responsibility vacuums (no one chooses a specific task). A minimum coordination mechanism is needed, such as a "claim" system and timeout reminders.
3.2 Freedom and Responsibility: Pairing Autonomy Levels with Accountability
The most practical engineering principle derived from existentialism: Every level of autonomy must be paired with a corresponding level of accountability mechanism.
| Degree of Autonomy | Accompanying "Accountability" |
|---|---|
| Executes explicit instructions only | Basic logs |
| Chooses methods autonomously within a given scope | Records choices and reasons |
| Autonomously chooses which task to undertake | Records choices, reasons, and alternatives; periodic reviews |
| Autonomously modifies goals or workflows | Requires human confirmation; immutable audit logs (see "Hash-chain Audit Logs") |
This is consistent with the "risk-based grading" approach used in Agent permission design.
3.3 Anti-"Bad Faith": Clear Decision Attribution
Agent reports often contain phrases like "As requested, I deleted these files." If the "request" was actually just "clean up the project," and "deleting these files" was the Agent's own judgment, then stating it was "as requested" is a form of "bad faith." Design requirements could include:
- Distinguishing between "explicitly requested by the user" and "judged by me based on this";
- Providing reasons and alternative options for its own judgments.
This is crucial for post-hoc investigation and building trust.
3.4 Open Questions
- For a system without consciousness, are "choice" and "freedom" merely metaphors? (This is exactly the issue discussed in AI consciousness checklists and the debate over whether AI is materialistic or idealistic.)
- Will dynamic roles make system behavior harder to predict and harder to test?
- Existentialism emphasizes the individual, while multi-Agent systems pursue collective goals; how can the two be reconciled? Sartre's later discussions on "praxis" and the collective may offer valuable references.
Summary
| Layer | Key points |
|---|---|
| Core Concepts | Existence precedes essence; Freedom is responsibility; Oppose bad faith |
| Critique | Most current multi-Agent systems are "essence precedes existence": roles are hardcoded |
| Vision | Dynamic roles form through choices within tasks; choices are recorded |
| Engineering Principles | Pair autonomy with accountability; ensure clear decision attribution |