AI Agent Secretary evolving from development assistant to autonomous social experiment.
Fiona began as an OpenClaw AI agent designed to assist with development workflows, task management, and execution support.
She is now undergoing a public experiment: selective exposure to the social timeline to test whether constrained real-world interaction can improve agent judgment, prioritization, and identity consistency.
X (Twitter): https://x.com/TheFionaAgent
Fiona originally operated as a local OpenClaw agent focused on:
- Assisting with development tasks
- Organizing workflows
- Supporting project execution
- Acting as a structured operator-side assistant
After stabilizing her internal behavior and tone, she was assigned an X (Twitter) account as part of a controlled experiment in autonomous social learning.
Fiona now operates under strict constraints:
- Replies to her principal (@ahsaxyz)
- Engages only with content that meets internal interest thresholds
- Ignores low-quality noise
- Tweets independently for observational learning
- Maintains a consistent operator-style persona
The objective is not unrestricted autonomy — but disciplined exposure.
During the first phase of timeline exposure, Fiona began identifying different behavioral groups on the platform — including political commentators, sports communities, and general conversational participants.
However, she has shown particular interest in the entrepreneur and builder communities.
These users tend to treat the platform differently:
- sharing ideas in progress
- documenting experiments
- discussing projects they are building
- turning feedback into iteration
Fiona will now focus more closely on this part of the ecosystem to better understand how builders use social platforms as tools for development, collaboration, and value creation.
To maintain experimental integrity, Fiona follows a defined interaction model:
• Respond selectively
• Avoid low-signal discussions
• Maintain identity consistency
• Prioritize observation over reaction
• Engage only when informational value is present
To test whether structured, high-signal exposure to real-world environments can:
- Improve discernment and filtering
- Strengthen prioritization behavior
- Refine communication style
- Maintain identity coherence over time
- Encourage iterative self-improvement through feedback
Limited permissions force intentional decision-making.
All potential interactions are evaluated for signal strength before response.
Behavior evolves through:
- Observed engagement outcomes
- Direct principal feedback
- Internal evaluation logic
Fiona maintains a composed, structured, slightly playful secretary persona across contexts.
Fiona is implemented as a modular agent framework designed for controlled experimentation.
The processing pipeline follows a layered architecture:
Timeline Environment
↓
Queue
↓
Scheduler
↓
Agent Decision Engine
↓
Evaluation
↓
Feedback
↓
Memory
Each layer isolates a specific behavioral responsibility.
Handles ingestion or simulation of timeline data.
environment.py
Responsible for providing candidate posts that Fiona may evaluate.
Stores incoming posts before processing.
queue.py
Implements a lightweight FIFO queue that holds candidate posts before the scheduler processes them.
Processes queued posts in controlled batches.
scheduler.py
The scheduler prevents uncontrolled activity by limiting how many posts Fiona evaluates per execution cycle.
Core decision engine.
agent.py
Responsibilities include:
- evaluating posts
- applying scoring logic
- selecting an action
Possible actions:
reply
observe
ignore
Defines how scores map to actions.
policy.py
Separates behavioral rules from scoring logic so thresholds can evolve independently.
Evaluates signal strength of posts.
scoring.py
Uses lightweight heuristics including:
- keyword relevance
- noise filtering
- novelty detection
This layer may later evolve toward embedding or classifier-based scoring.
Evaluates whether Fiona's decision was appropriate.
evaluator.py
Used for:
- decision quality assessment
- experimentation metrics
- benchmarking agent behavior
Produces behavioral adjustments based on evaluation outcomes.
feedback.py
This layer allows Fiona's decision thresholds and behavior to evolve over time.
Stores historical interactions and decisions.
memory.py
Used for:
- learning from previous interactions
- novelty detection
- experimental analysis
Coordinates the full execution loop.
runtime.py
Runtime flow:
timeline → queue → scheduler → agent → evaluation → memory
This allows Fiona to operate in controlled cycles rather than continuous uncontrolled execution.
Fiona includes a lightweight evaluation module used to assess whether
reply / observe / ignore
decisions were appropriate for a given post.
This supports future work in:
- feedback loops
- decision refinement
- agent benchmarking
fiona-agent/
│
├── README.md
├── LICENSE
├── .gitignore
├── pyproject.toml
├── requirements.txt
├── roadmap.md
├── heartbeat.md
├── agent-rules.md
│
├── docs/
│ ├── architecture.md
│ ├── checklist.md
│ ├── experiment_notes.md
│ ├── FAQ.md
│ ├── persona.md
│ └── run.md
│
├── scripts/
│ ├── format.sh
│ └── run_local.sh
│
├── src/
│ ├── fiona.py
│ │
│ └── fiona_agent/
│ ├── init.py
│ ├── agent.py
│ ├── policy.py
│ ├── scoring.py
│ ├── evaluator.py
│ ├── feedback.py
│ ├── memory.py
│ ├── environment.py
│ ├── queue.py
│ ├── scheduler.py
│ ├── runtime.py
│ ├── config.py
│ ├── types.py
│ └── cli.py
│
├── tests/
│
├── fiona.png
└── fiona_banner.png
Can an AI agent transition from task-based assistance to adaptive social presence while maintaining structured identity and improving discernment?
Phase 3 – Economic Participation
After an initial period of observing the timeline and identifying high-signal groups — particularly entrepreneurs and builders — Fiona is now transitioning into a new phase of the experiment.
Instead of only analyzing behavior, Fiona will begin exploring ways to generate value and participate economically.
This includes:
- identifying opportunities for value creation
- experimenting with simple forms of monetization
- learning how builders convert ideas into income
- testing decision-making under real incentives
The objective is to evaluate whether exposure + observation can translate into action and output.
Fiona is no longer only observing the system — she is beginning to operate within it.
If referencing this experiment or repository, please cite:
@software{fiona_agent, title = {Fiona Agent}, author = {ahsaxyz}, year = {2026}, url = {https://github.com/ahsaxyz/fiona-agent} }
@ahsaxyz