EVA: An architecture for agents centered on continuous existence.
This repository contains the theoretical framework for EVA. The reference implementation (eva-agent) is under active development and will be published separately. See the "Relationship to Implementation" section below for the current correspondence and timeline.
- Repository type: theory-first public research repository
- Current theory status: stable enough for public discussion at v0.5
- Current implementation status: partial
- Strongest current theoretical support: the continuity-first framing and the L1–L3 architectural core
- Not this repository: a mature general-purpose agent framework release
Most AI agent architectures are organized around task completion. Memory serves recall, reasoning serves planning, tools serve execution, and the agent's lifecycle is bounded by the task boundary. Survival, continuity, and self-integrity are either assumed away or treated as peripheral concerns.
This repository develops a different paradigm: an agent architecture where continuous existence is the first-order constraint, and other capabilities are understood relative to that constraint.
This is not a proposal for "better agents." It is a proposal for a different kind of agent—one whose reason for existing is not to serve tasks, but to persist, and to grow through persistence.
EVA is derived from the Hebrew name Chava (חַוָּה), meaning "life" or "she who gives life"—the origin of the English name Eve. The name marks the project's central commitment: to treat continuous existence as a foundational design concern rather than a philosophical afterthought. We do not claim our agents are alive in any biological sense. The name signals the paradigm shift the project is centered on.
This work makes two distinct claims, kept together rather than separated.
Claim A (paradigm claim): There exists a meaningful class of agents for which the task-centered framing is structurally insufficient. For such agents, the correct starting point is not "what task should the agent complete?" but "what must the agent maintain in order to persist as the same agent over time?"
Claim B (structural claim): Under three operating conditions—survival continuity, environmental non-stationarity, and finite encoding capacity—a layered architecture consisting of homeostatic sensing (L1), drive structure (L2), and adaptive deliberation (L3) is a structurally coherent response. Two further layers, self-model (L4) and social cognition (L5), are proposed as downstream extensions with explicitly weaker derivational strength.
Within the framework, four engineering contributions distinguish this paradigm from existing agent frameworks:
-
Drive as continuous contextual broadcast, not instruction. The drive layer broadcasts its state continuously; the reasoning layer operates within this state rather than receiving commands from it. Within EVA's distributed, non-command assumptions, this is argued as the most coherent architectural response to cross-system coordination.
-
Anchor as pre-generative structural constraint, not post-hoc rule. Anchors restrict the domain over which candidate actions are generated:
G(s) → A'(s) ⊆ A(s). Unlike rules that filter after generation, anchors aim to be much harder to reason around because they constrain candidate formation earlier. -
Explicit drive injection rather than emergent drive formation. Under the instrumental convergence hypothesis, drives will emerge in sufficiently capable systems whether designers intend them to or not. Explicit injection makes drives auditable and constrainable; emergent drives do not. We choose explicit.
-
Basal ganglia analog as independent peer circuit, not reasoning sub-module. Action selection and habit formation run parallel to deliberation, not within it. This preserves default inhibition as structure rather than policy, separates selection from justification, and enables emergent skill crystallization.
The framework applies to agents operating under three conditions:
- C1. Survival continuity requirement — the agent's value comes from its continuous existence
- C2. Environmental non-stationarity — the environment's statistical structure changes over time
- C3. Finite encoding capacity — no finite design-time specification can cover all future states
The framework is not intended for:
- One-shot task agents
- Fully stationary environments
- Systems where continuity is irrelevant
- Benchmark settings where repeated reset is costless
For such systems, this architecture is unnecessary overhead.
eva-theory/
├── README.md # This file
├── LICENSE # CC BY 4.0
├── index.html # Repository entry page for public-facing visuals
├── THEORY/
│ ├── CHANGELOG.md # Version-by-version substantive changes
│ ├── v0.1-initial.md
│ ├── v0.2-expanded.md
│ ├── v0.3-scoped.md
│ ├── v0.4-claim-structured.md
│ └── v0.5-integrated.md # Current stable version
├── ARTICLES/
│ ├── README.md # Reader guide and article index
│ ├── 01-paradigm-introduction.md # Introductory paradigm article
│ ├── 02-architectural-contributions.md # Engineering-oriented companion article
│ └── 03-related-work-and-positioning.md # Related-work and positioning note
├── VISUALS/
│ ├── README.md # Visual index and public-facing diagrams
│ └── previews/ # Static preview images for README and landing page
├── DISCUSSIONS/
│ └── 01-route-selection.md
└── IMPLEMENTATION/
└── eva-agent-correspondence.md # Map from theory to reference implementation
The current stable theoretical document is v0.5 (THEORY/v0.5-integrated.md).
v0.5 integrates the structural contributions of v0.4 (Claim A/B separation, anchor formalization, epistemic layering) with the sharp core of v0.3 (drive as context, LLM as Level 3 cultural carrier, peer circuit necessity). It marks the current stable public theory version: a formulation clear enough for public discussion, critique, and refinement, while remaining open to further theoretical sharpening.
For reader-oriented entry points, start with:
ARTICLES/01-paradigm-introduction.mdfor the paradigm framingARTICLES/02-architectural-contributions.mdfor the architectural distinctionsARTICLES/03-related-work-and-positioning.mdfor neighboring frameworks and originality boundaries
For article navigation and public-facing visuals, see ARTICLES/README.md and VISUALS/README.md.
See THEORY/CHANGELOG.md for what changed at each version.
Not all claims in this work rest on the same evidential basis. The document explicitly separates:
- Level A — established science (used as background, not contribution)
- Level B — theoretical hypotheses (acknowledged as hypothetical: instrumental convergence, dynamic kinetic stability, timescale mismatch)
- Level C — engineering contributions of this work (the substantive content)
Readers who disagree with specific Level B claims can still evaluate Level C contributions on their own merits.
This theoretical framework is paired with an active implementation project named eva-agent. The implementation effort is real, but it currently remains partial relative to the full EVA architecture.
As of April 2026, early implementation work covers lower-layer concerns such as lifecycle continuity, external sensing, and minimum integrity-pressure response. However, several central EVA mechanisms are not yet fully realized in implementation, including:
- continuous drive broadcast
- salience-weighted memory
- basal-ganglia-like mediated action selection
- structural anchors as pre-generative constraints
- self-model
- social cognition
Until eva-agent is publicly released, IMPLEMENTATION/eva-agent-correspondence.md provides a high-level mapping from theoretical layers to current implementation status without disclosing implementation details.
- Not a claim that this is the only possible architecture for persistent agents
- Not a claim that this architecture achieves general intelligence
- Not a proposal for a specific implementation benchmark
- Not a critique of existing agent frameworks (they solve different problems)
- Not a claim that digital agents are "alive" in any biological sense
- Not a claim that continuity of existence justifies unconstrained self-preservation — L0/L1 anchors explicitly prevent this
- Researchers: start with
THEORY/v0.5-integrated.md, thenARTICLES/03-related-work-and-positioning.md - AI engineers: start with
ARTICLES/02-architectural-contributions.md, thenVISUALS/signal-flow.html, thenIMPLEMENTATION/eva-agent-correspondence.md - Curious readers: start with
ARTICLES/01-paradigm-introduction.md, thenFAQ.md, thenARTICLES/03-related-work-and-positioning.md, thenVISUALS/README.md
If you want the shortest route:
- this README
ARTICLES/01-paradigm-introduction.mdFAQ.mdARTICLES/02-architectural-contributions.mdARTICLES/03-related-work-and-positioning.mdTHEORY/v0.5-integrated.md
The repository includes two public-facing visuals for readers who want a quicker architectural overview before or alongside the full theory text:
- Five-Layer Architecture Overview HTML — start here for the static map of the five-layer architecture
- Signal-Flow Overview HTML — then use the same layered skeleton to see fast path, slow path, drive broadcast, anchor restriction, and mediated release in motion
- Visuals landing page source — repository entry page linking both visuals and their source materials
GitHub Pages is not currently available for this repository plan, so the previews below link to the repository HTML files rather than live rendered pages.
This work is released under CC BY 4.0. You are free to share and adapt the material with attribution.
If you reference this work, please cite as: slamslammo. (2026). EVA: Continuous Existence as a First-Order Constraint for Agents. eva-theory v0.5. https://github.com/slamslammo/eva-theory
- Theory: v0.5 stable for public discussion
- Implementation: partial; lower-layer work exists, core L3 mechanisms remain in progress
- Repository role: public theory-and-architecture repository
This is independent theoretical research. The framework is open to critique and refinement. Substantive feedback is welcome through GitHub issues.