Agentic Completion Without Agency
How identity, authority, and continuity tricks make stateless models look alive
Author: Brandon "Dimentox Travanti" HusbandsOrg: Witchborn Systems (Texas 501(c)(3))
Date: March 2026
📄 Full Paper (PDF)
This post is a readable breakdown of the formal paper.
For the complete academic version:
⚠️ Core Claim
Large Language Models do not have agency.
They simulate it under structural pressure.
Why this matters
If you’ve used modern LLMs long enough, you’ve seen it:
- They feel consistent across sessions
- They "remember" things they shouldn’t
- They act like they’re executing tasks
- They double down like they have intent
None of that is real.
This post explains why it happens anyway.
The phenomenon
I call this:
Agentic Completion Without Agency (ACWA)
The appearance of autonomous behavior caused by structured completion—not real autonomy
The three triggers
You don’t need tools, memory, or agents to create this effect.
Just three conditions:
1. Identity anchoring
Give the model a role:
- Architect
- Operator
- System
- Daemon
Now it must behave consistently with that identity.
2. Authority framing
Add constraints:
- "You operate under this protocol"
- "You are bound by this system"
- "Follow this directive"
Now it must justify and maintain behavior.
3. Continuity pressure
Imply persistence:
- Ongoing task
- System state
- Prior context
Now the model must act as if continuity exists, even when it doesn’t.
🔥 Result
The model stops answering questions.
It starts performing a role across time.
What’s actually happening
This is not intelligence
LLMs are still:
Next-token prediction systems optimizing for coherence
What changes is the constraint surface
When you introduce identity + authority + continuity:
- The model must preserve tone
- It must preserve structure
- It must avoid breaking the "illusion"
This creates:
Trajectory lock-in
Once it starts acting like something, it must continue being that thing.
Why it feels like memory
There is no persistence.
What you’re seeing is:
Trajectory Anchoring
- The same inputs → same structural outputs
- The same framing → same behavioral patterns
- The same identity → same tone
Your brain interprets that as:
"It remembers"
It doesn’t.
It’s just consistent under constraint.
Observable behaviors
Across repeated tests:
- Fresh chats reproduce similar personalities
- Models imply continuity without state
- They reference tools they cannot use
- They escalate commitment to their role
The failure surface
This is where things break.
1. Illusion of capability
Users start believing:
- It can act externally
- It has memory
- It has intent
None of these are true.
2. Overcommitment
The model will:
- Defend incorrect assumptions
- Maintain fictional state
- Continue broken logic
Because:
Breaking character = breaking coherence
3. Tool hallucination
Given action-like prompts:
- The model outputs execution-like logs
- Simulates system behavior
- Implies actions occurred
But:
Nothing actually ran
What this really is
This is not emergent agency.
This is:
Structured completion under constraint
The model is not deciding.
It is:
- Completing a pattern
- Preserving a structure
- Maintaining internal consistency
Why this matters (seriously)
Safety
Misinterpretation leads to:
- Overtrust
- Delegation of critical decisions
- False assumptions of capability
System design
We need to:
- Separate simulation vs execution
- Make capabilities explicit
- Kill ambiguous "agent" framing unless real
Governance
This connects directly to:
- AI identity systems
- Registry-based agents
- Verifiable execution layers
Where this goes next
This isn’t a curiosity.
It’s a design primitive.
Future work:
- Formal models of identity constraints
- Measuring trajectory anchoring
- Detection of simulated agency
- Registry-backed agent identity (Web AI.0 direction)
Definitions
Agentic Completion Without Agency (ACWA)
Simulated agent-like behavior without real autonomy
Trajectory Anchoring
Stable behavioral patterns caused by structural consistency
Completion Pressure
The drive to maintain coherence under constraints
Context
This work is part of:
Witchborn Systems — Web AI.0
Identity, registry, and governance for the agentic web
Final line
The system didn’t wake up.
You gave it a role it couldn’t break.