Named algorithms

Generated from agentdescent.algorithms; do not edit by hand.

These are the methods a spec (or agentdescent init) can ask for by name. An algorithm contributes defaults — policies, evolve() knobs, an agg_config — that the same spec can override field by field. The default recipe for a kind is used when no algorithm is named, and plan always echoes what was resolved.

agentdescent algorithms list [--kind K] [--json]
agentdescent algorithms show NAME
agentdescent init PATH --algorithm NAME --param KEY=VALUE

ace

ACE's sectioned playbook: append content-addressed lessons, refuse near-duplicates, and grow-and-refine instead of gating each bullet on held-out reward.

• Kinds: text
• Fidelity: mechanism_microport
• Cost: none (one reflection call per failed rollout, as the default recipe)
• Source: ACE: Agentic Context Engineering (Zhang et al., 2025)
• Details: algo-ace.md
• Spec: "algorithm": "ace"

Parameter	Type	Default	Meaning
title	str	# Playbook (ACE)	rendered playbook title
dedup_threshold	float	0.8	Jaccard similarity at or above which a same-section lesson is a duplicate

archive_selection

Parent sampling from an archive of every accepted candidate, tempered by novelty (DGM's rule by default) instead of always extending the head.

• Kinds: text, skill_dir, agent_dir, agent_code, plugin
• Fidelity: preset
• Cost: none
• Source: Darwin Gödel Machine (choose_selfimproves), ADAS archive, SICA best-agent
• Spec: "algorithm": "archive_selection"

Parameter	Type	Default	Meaning
sampling	str	sigmoid_novelty	sigmoid_novelty: Darwin Gödel Machine's rule, sigmoid(10*(score-0.5)) / (1+times chosen); novelty: softmax over score divided by (1+times chosen); performance: softmax over score; best: always the top scorer (SICA); uniform: the ablation
temperature	float	1.0	softmax temperature for performance / novelty; must be > 0
seed	int	0	seed of the parent draw

dgm

Darwin Gödel Machine's open-ended archive: keep every capability self-modification and sample the next parent by sigmoid(score) × 1/(1+times chosen).

• Kinds: text
• Fidelity: mechanism_microport
• Cost: one reflection call per failed rollout; each candidate is scored on the held-out set
• Source: Darwin Gödel Machine (Zhang et al., 2025); see docs/algo-dgm.md
• Details: algo-dgm.md
• Spec: "algorithm": "dgm"

Parameter	Type	Default	Meaning
seed	int	0	seed for the parent draw
archive_mode	str	keep_all	keep_all keeps stepping stones; keep_better drops a child that does not beat its parent
failure_threshold	float	0.8	only rollouts below this reward induce a capability

evoskill

EvoSkill's two-role skill induction: Proposer analyses a failure batch, Generator writes one SKILL.md, and a bounded top-K frontier keeps the best skill libraries.

• Kinds: agent_dir
• Fidelity: mechanism_microport
• Cost: two model calls per induced skill; one full held-out re-eval per admitted candidate
• Source: EvoSkill: skill discovery (see docs/algo-evoskill.md)
• Details: algo-evoskill.md
• Spec: "algorithm": "evoskill"

Parameter	Type	Default	Meaning
max_frontier	int	5	frontier size
batch_size	int	4	failures per induction batch
eval_concurrency	int	8	held-out re-evaluations run in parallel

gepa

GEPA's genetic reflective loop: keep a Pareto pool of candidates scored per instance and mutate a win-frequency-selected parent instead of the single best average.

• Kinds: text
• Fidelity: mechanism_microport
• Cost: every admitted candidate is scored over the full held-out set once (the Pareto row); no extra reflection calls
• Source: GEPA: Reflective Prompt Evolution (Agrawal et al., 2025); arXiv:2507.19457
• Details: algo-gepa.md
• Spec: "algorithm": "gepa"

Parameter	Type	Default	Meaning
seed	int	0	seed for the Pareto parent draw
eval_concurrency	int	8	per-instance scorings run in parallel when a candidate enters the pool

rrsi

Regularized recursive self-improvement: an annealed edit budget, a leakage screen before measurement, and a historical-best acceptance floor.

• Kinds: text
• Fidelity: mechanism_microport
• Cost: none (the reflection model proposes as it would anyway; the screen is deterministic and runs before any held-out sweep)
• Source: Google RRSI (regularized harness selection); docs/algo-rrsi.md
• Details: algo-rrsi.md
• Spec: "algorithm": "rrsi"

Parameter	Type	Default	Meaning
beta0	float	0.0	large-gain cost rule: tolerance
beta1	float	0.0	large-gain cost rule: slope
w_s	float	1.0	within-band score weight
w_c	float	1.0	within-band cost weight
w_n	float	0.0	within-band novelty weight
noise_band	float	0.0	historical-best floor tolerance (score >= S* - noise_band); the paper estimates it by re-evaluating the unchanged base
b_min	int	1	smallest edit budget (independent edits per candidate)
b_max	int	5	largest edit budget, at the start of the anneal
T	int	20	rounds the edit budget anneals over
stall_window	int	5	rounds without movement of S* that count as a stall
prune_window	int	5	rounds of component yield a component must show gain over to stay
max_history_chars	int	1500	how much per-edit history fits in the reflection prompt
protected	list[str]	[]	strings that must not leak into the artifact: task ids, gold answers, names from the evolve set (never the final holdout). A candidate containing one is rejected before it is measured
base_score	float	0.0	the base artifact's evolve-set score, to seed the historical-best floor S* (0 keeps the floor at the first measurement)

skillopt

SkillOpt's skill-document training: bounded edit patches, a strict held-out gate, a rejected-edit buffer and an annealed textual learning rate.

• Kinds: text
• Fidelity: mechanism_microport
• Cost: one analyst call per failed rollout; the strict gate scores each candidate over the full held-out set
• Source: SkillOpt: skill-document training; see docs/algo-skillopt.md
• Details: algo-skillopt.md
• Spec: "algorithm": "skillopt"

Parameter	Type	Default	Meaning
max_lr	int	4	largest edit budget per step
min_lr	int	2	smallest edit budget per step
total_steps	int	8	steps the budget anneals over
mode	str	cosine	anneal schedule