Layerd AI Agent Reforge

Layerd AI Agent Reforge is a Python library for the outer loop of agent development: generating program variants, running them under constraints, scoring them with evaluators, archiving results, optionally gating promotion on human approval, and evolving an editable reforge procedure (the improvement pipeline itself can change). It is meant to sit next to runtimes like LangGraph—your graph still executes tasks; Reforge versions and improves the unified program (prompts, tools, graph hints, evaluator wiring) that you compile into that runtime.

Spelling: the product short name is Reforge (with a final e). The PyPI package remains layai-agent-reforge; the import path is layai_reforge.

Research concepts are inspired by Meta’s HyperAgents / DGM-H (HyperAgents publication). This project is not affiliated with or endorsed by Meta.

---

Table of contents

• What this package is (and is not)
• Core concepts
• How the improvement loop works
• Installation
• How to use it
• Program patches
• Sandboxing
• Evaluators
• Human approval
• Reforge procedure engine
• CLI (layai-reforge)
• Examples
• Documentation
• Developing & testing
• License

---

What this package is (and is not)

Reforge is	Reforge is not
A framework for variants → runs → evaluation → archive → optional promotion	A replacement for LangGraph, AutoGen, or your chat UI
A place to define UnifiedProgram (task + reforge agent + reforge procedure) as data	An LLM hosting service
Sandbox helpers (SandboxRunner) for subprocess-style runs with defaults like no network	A full MLOps or deployment platform
HumanGate hooks before treating a variant as canonical	A built-in web dashboard (you can add one on top of the API)

Promotion produces an in-memory UnifiedProgram snapshot (ImprovementLoopResult.promoted_program). Deploying that snapshot to production is your application’s responsibility.

---

Core concepts

• UnifiedProgram — One structured artifact: TaskAgentSpec (system prompt, tools, graph builder ref, graph_config, limits), ReforgeAgentSpec (prompts for the reforge agent that proposes program patches), and ReforgeProcedureSpec (declarative pipeline steps + evaluator_ids used by the improvement loop when scoring runs).
• Variant — A set of ProgramPatchOp entries applied to a parent program to form a child candidate.
• VariantGenerator — Builds variants (e.g. paraphrase prompt, tool subsets, LLM-proposed patches if you supply an llm_fn).
• ImprovementLoop — run_generation drives: materialize variants → your run_artifact_fn produces RunArtifact → evaluators (via EvaluatorRegistry) → PromotionPolicy → optional HumanGate → optional promoted_program; every candidate is also archived via ArchiveStore.
• ArchiveStore — Persists stepping-stone programs and scores (e.g. SqliteArchiveStore; optional Postgres extra).
• ReforgeSession — Convenience wrapper: default sandbox, default MathGradingEvaluator registration, ledger/reforge-memory/audit wiring, and helpers like run_improvement_generation(loop, ...) and run_reforge_pipeline().

---

How the improvement loop works

At a high level:

1. You supply a base UnifiedProgram and a variant factory (list of Variant).
2. For each variant, the library materializes a child program (apply_patches), then calls your run_artifact_fn(child, variant) to obtain a RunArtifact (stdout/stderr, exit code, optional extra).
3. Evaluators named in child.reforge_procedure.evaluator_ids are resolved and composed; an aggregate score is computed (CompositeEvaluator).
4. The best-scoring variant is a candidate for promotion if PromotionPolicy.should_promote passes (default: minimum aggregate score 0.8 and all sub-evaluators must pass when require_all_evaluators_pass is true).
5. If a HumanGate is configured, APPROVE is required before promoted_program is set; otherwise the gate can auto-approve (useful for tests).
6. Each variant’s run is recorded as an ArchiveEntry in the store.

flowchart TD
  UP[Base UnifiedProgram] --> VF[variant_factory]
  VF --> MAT[materialize via patches]
  MAT --> RUN[run_artifact_fn → RunArtifact]
  RUN --> EV[CompositeEvaluator / registry]
  EV --> PP[PromotionPolicy]
  PP --> HG{HumanGate?}
  HG -->|approve| PROM[promoted_program]
  HG -->|reject / no promote| NOP[no promotion]
  RUN --> AR[ArchiveStore.add_entry]

The ReforgeProcedureEngine is a separate path: it walks ReforgeProcedureSpec.steps (retrieve archive, propose patch, static lint, etc.) and is typically used for outer-loop edits and experimentation, not as a substitute for wiring ImprovementLoop for your task runs.

---

Installation

pip install layai-agent-reforge

Optional extras:

Extra	Purpose
langgraph	LangGraph adapters (layai_reforge.adapters.langgraph)
postgres	Postgres-backed archive
dev	pytest, pytest-asyncio for contributors

pip install "layai-agent-reforge[langgraph,dev]"

---

How to use it

1. Define a program and archive

from pathlib import Path

from layai_reforge import SqliteArchiveStore, UnifiedProgram
from layai_reforge.models.program import ReforgeProcedureSpec, TaskAgentSpec

program = UnifiedProgram(
    task=TaskAgentSpec(system_prompt="You are a careful assistant."),
    reforge_procedure=ReforgeProcedureSpec(evaluator_ids=["math_grade"]),
)
archive = SqliteArchiveStore(Path(".reforge/archive.sqlite"))

2. Wire an improvement loop

You provide variant_factory(base) -> list[Variant] and run_artifact_fn(child, variant) -> RunArtifact. The artifact is what evaluators score (e.g. stdout for MathGradingEvaluator).

from layai_reforge.evaluators.math_eval import MathGradingEvaluator
from layai_reforge.evaluators.registry import EvaluatorRegistry
from layai_reforge.loop.improvement import ImprovementLoop
from layai_reforge.loop.variants import VariantGenerator
from layai_reforge.models.artifacts import RunArtifact
from layai_reforge.sandbox.runner import SandboxConfig, SandboxRunner

sandbox = SandboxRunner(SandboxConfig(workspace_root=Path.cwd()))
registry = EvaluatorRegistry()
registry.register(MathGradingEvaluator(golden_answer="42"))

vg = VariantGenerator(seed=0)
variants = [vg.paraphrase_prompt_variant(program)]

def run_artifact_fn(child, variant):
    return RunArtifact(
        run_id="r1",
        variant_id=variant.id,
        stdout="42",
        success=True,
    )

loop = ImprovementLoop(archive=archive, sandbox=sandbox, registry=registry)
result = loop.run_generation(
    program,
    variant_factory=lambda p: variants,
    run_artifact_fn=run_artifact_fn,
)
# result.promoted_program may be None (policy / human gate / scores)

3. Optional: ReforgeSession

from layai_reforge import ReforgeSession

session = ReforgeSession(program=program, archive=archive)
result = session.run_improvement_generation(loop, variant_factory=lambda p: variants, run_artifact_fn=run_artifact_fn)

Register any custom evaluators on session.registry (or build your own EvaluatorRegistry for ImprovementLoop) so IDs in reforge_procedure.evaluator_ids resolve.

---

Program patches

Patches are validated ProgramPatchOp values applied by apply_patches (see src/layai_reforge/patches.py). Supported operations include:

Operation	Role
set_system_prompt	Replace task system prompt
add_tool / remove_tool	Mutate ToolDescriptor list
set_graph_config_key	Set a key under task.graph_config (path-safe)
set_reforge_patch_prompt / set_reforge_procedure_patch_prompt	Reforge agent prompts
replace_reforge_procedure_steps	Replace ReforgeProcedureStep list
set_reforge_procedure_evaluators	Set evaluator IDs for scoring

Unknown ops are rejected.

---

Sandboxing

SandboxConfig defaults include a wall-clock timeout, allow_network=False, and an env_allowlist for subprocess environments. SandboxRunner.run_command executes under these constraints and returns RunArtifact-oriented results; see src/layai_reforge/sandbox/runner.py for backends (subprocess default; optional Docker).

---

Evaluators

Built-in examples live under layai_reforge.evaluators: e.g. MathGradingEvaluator, coding/paper/robotics helpers. Register them by id so they match ReforgeProcedureSpec.evaluator_ids. CompositeEvaluator merges multiple evaluators and writes an aggregate metric (configurable key).

---

Human approval

HumanGate can auto-approve/reject (tests, batch jobs) or take an async on_request(variant, report) callback for interactive UIs or ticket systems. Until approval, ImprovementLoopResult.promoted_program stays None when the gate rejects.

---

Reforge procedure engine

ReforgeProcedureEngine.run(ctx) executes ReforgeProcedureSpec.steps: e.g. retrieve_archive, propose_patch (if you pass propose_patch_fn), static_lint, simulate_rollout (hook placeholder), aggregate_scores, rollback, summarize_reforge_memory. Use this to experiment with how the reforge side proposes changes; combine with ReforgeSession.run_reforge_pipeline for a single entrypoint.

Schema: saved programs use schema_version "2" with fields reforge_agent and reforge_procedure. Legacy files that used meta / meta_procedure (v1) are still accepted by load_program / UnifiedProgram.model_validate via a migration hook.

---

CLI (layai-reforge)

layai-reforge --help

Useful commands:

Command	Purpose
init <path>	Write a starter program JSON/YAML
archive-list --db …	List SQLite archive entries
export / import	Export or merge archive bundle JSON
eval-once <program>	Load program and print fingerprint JSON

Subcommands like promote, run-loop, and replay currently print guidance: promotion and the full outer loop are intended to be driven from Python (ImprovementLoop, HumanGate, your telemetry).

---

Examples

Path	What it shows
examples/minimal_loop.py	Minimal variant → artifact → MathGradingEvaluator → archive
examples/cross_domain_transfer.py	Transfer / domain tagging patterns
examples/hypothetical_status_agent.py	End-to-end stub “health check” scenario (simulated stdout from prompt inspection); run with python examples/hypothetical_status_agent.py from the repo (or install the package)

From a checkout:

pip install -e ".[dev]"
pytest
python examples/hypothetical_status_agent.py

---

Documentation

• docs/concepts.md — Maps HyperAgents-style ideas to Reforge modules.
• memory-bank/project-context.md — Maintainer/agent notes (layout, integration, publishing checklist).

---

Developing & testing

git clone <your-repo-url>
cd layai-agent-reforge
python3 -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
pytest

---

License

MIT — see the LICENSE file in the repository root.