FALM: Fitness-Aware Agent Lifecycle Management

Recursive Fitness Evaluation and Death as a Convergence Primitive in Self-Regenerating Agentic Systems

Author: Anurag Krishna Sharma
Affiliation: Advanced Data Processing Research Institute (ADRIN), Department of Space, Government of India
Target Venue: NeurIPS 2026 LLM Agents Workshop (8-page limit)

---

What is FALM?

FALM is a framework for self-regenerating agentic systems where:

1. Agents spawn agents — elite agents reproduce via crossover + mutation of cognitive gene vectors
2. The fitness metric co-evolves — the Anurag Index weight vector Φ = (α, β, γ, δ) is itself evaluated and updated by a meta-level index AI(N+1)
3. Death is a design primitive — agents that detect their own sub-optimality trigger graceful termination, compressing learned state into offspring initialisation (biological apoptosis analogy)

The central novelty: the metric that selects agents is itself selected for.

---

The Anurag Index

Level N — agent fitness

AI(N)(a_i) = α · f(g_i) + β · ln(gen_i + 1) + γ · τ_i + δ · AI(N)(a_parent)

Term	Symbol	Meaning
Intrinsic capability	α · f(g_i)	weighted aggregation over 8-dim gene vector
Generational depth	β · ln(gen+1)	log-scaled lineage bonus
Task success	γ · τ_i	empirical performance across task sample
Hereditary bonus	δ · AI(N)(a_parent)	Lamarckian head-start from parent

Constraint: α + β + γ + δ = 1, all ≥ 0. Index ∈ [0, 10] after scaling.

Gene vector dimensions: PLAN · EXEC · MEM · TOOL · LEARN · COMM · SAFE · SPEED

Level N+1 — meta-fitness (evaluates the evaluator)

AI(N+1)(Φ) = E_{T ~ D} [ Spearman_Corr( AI(N)(a; Φ),  TruePerf(a, T) ) ]

Weight update: Φ ← proj_simplex( Φ + η · ∇_Φ AI(N+1)(Φ) )

Death condition D*

D*(a_i) = 1[ AI(N)(a_i) < θ · AI(N)(a_best) ]  ∧  1[ ∇_t AI(N)(a_i) < 0 ]

On D*: agent compresses state bundle S_i → child gene initialisation
g_child = crossover(g_parent, decode(S_i)) + ε, ε ~ N(0, σ²)

---

Repo Structure

falm/
├── README.md                  ← this file
├── paper/
│   ├── outline.md             ← full section-by-section paper outline
│   ├── theory.md              ← formal proofs (Conjecture 1, Proposition 1)
│   └── related_work.md        ← annotated bibliography
├── falm/
│   ├── __init__.py
│   ├── agent.py               ← Agent class (gene vector, lifecycle)
│   ├── index.py               ← Anurag Index AI(N) computation
│   ├── meta_index.py          ← Meta-index AI(N+1), weight update
│   ├── population.py          ← Population dynamics (reproduce, prune, D*)
│   └── benchmarks/
│       ├── reasoning.py       ← GSM8K, StrategyQA wrappers
│       ├── coding.py          ← HumanEval, MBPP wrappers
│       └── retrieval.py       ← NaturalQuestions, TriviaQA wrappers
├── experiments/
│   ├── run_baselines.py       ← Fixed-α, Fixed-uniform, Random-spawn
│   ├── run_falm.py            ← Full FALM experiment
│   ├── ablations.py           ← A1–A4 ablation suite
│   └── configs/
│       ├── default.yaml
│       └── ablation_*.yaml
├── results/
│   └── .gitkeep
├── notebooks/
│   └── visualise_results.ipynb
├── requirements.txt
└── pyproject.toml

---

Paper Outline

See paper/outline.md for the full section-by-section outline.

Sections:

1. Introduction
2. Related Work
3. The FALM Framework (formal definitions)
4. Theoretical Analysis (proofs)
5. Experiments
6. Discussion
7. Conclusion + Appendices

---

Theoretical Results

Conjecture 1 — Fitness convergence

Under regularity conditions on f and task distribution D:

• Population-average E[AI(N)(a; Φ_t)] is non-decreasing in expectation
• Weight trajectory Φ_t converges to a local maximum of AI(N+1) in O(T · log T) generations

Proof sketch in paper/theory.md

Proposition 1 — Death accelerates convergence

Under D* with threshold θ and window w, the expected number of generations to reach target E[AI(N)] ≥ φ* is strictly less than under implicit pruning alone.

Proof by Markov chain comparison — see paper/theory.md

---

Experiments

Benchmarks

Suite	Tasks	Metric
Reasoning	GSM8K, StrategyQA	Accuracy
Code synthesis	HumanEval, MBPP	pass@1
Retrieval	NaturalQuestions, TriviaQA	Exact match

Baselines

Baseline	Description
Fixed-α	AI(N) with α=1, β=γ=δ=0
Fixed-uniform	AI(N) with α=β=γ=δ=0.25
No-death	FALM without D*
Random-spawn	No fitness selection
FALM (ours)	Full framework

Key ablations

• A1 — Co-evolving Φ vs. fixed Φ
• A2 — D* death vs. implicit pruning
• A3 — Hereditary δ term vs. no hereditary term
• A4 — Sensitivity to death threshold θ

Default hyperparameters

population_capacity: 50
elite_fraction: 0.4
mutation_sigma: 0.15
death_threshold: 0.6
death_window: 5
meta_lr: 0.01
gene_dims: 8

---

Quickstart

git clone https://github.com/akrishnash/falm
cd falm
pip install -e .

# run full FALM experiment
python experiments/run_falm.py --config experiments/configs/default.yaml

# run ablations
python experiments/ablations.py --suite A1 A2 A3 A4

# visualise
jupyter notebook notebooks/visualise_results.ipynb

---

Status

• Conceptual framework
• Mathematical definitions
• Paper outline
• Python implementation (in progress)
• Benchmark wrappers
• Experiments
• Paper write-up

---

Citation

@article{sharma2026falm,
  title   = {Fitness-Aware Agent Lifecycle Management: Recursive Fitness Evaluation
             and Death as a Convergence Primitive in Self-Regenerating Agentic Systems},
  author  = {Sharma, Anurag Krishna},
  journal = {NeurIPS 2026 Workshop on LLM Agents},
  year    = {2026}
}