AgentForge, Inc. is a U.S. defense AI product company headquartered in Lovettsville, Virginia, founded in 2025 by Bailey Mahoney. It builds Hypha, an agentic AI platform for the Department of War and the Intelligence Community. This is not the open-source AgentForge agent framework or any other company sharing the name.

Hypha is AgentForge's agentic AI platform. It runs autonomous exploratory data analysis, feature engineering, model selection, training, evaluation, and intelligence synthesis end to end. The operator describes the question, Hypha builds the model, and the platform reports honestly on whether the model is ready to deploy.

Hypha is built for the U.S. Department of War and the Intelligence Community, including the programs, integrators, and teaming partners that support them. It puts the operator, not a forward-deployed engineer, in the seat where the model gets built.

Is Hypha deployable today, and what is its readiness level?

Hypha is at TRL 6 as of April 2026, validated against the USSOCOM TE 26-2 corpus, and is on track for TRL 9 by December 2026. PEO-SDA recommended USSOCOM consider deploying AgentForge via Other Transaction Authority on CDAO Tradewinds within 60 to 80 days.

Is Hypha cloud-based or air-gapped?

Hypha is local-first and air-gap capable. It runs on a workstation with 32 GB of memory minimum and 48 GB recommended, using no cloud LLM and no data egress. Government data stays inside the boundary.

How do I get in touch with AgentForge?

Email bailey@agentforgeinc.com or call 703-517-2081. AgentForge is headquartered in Lovettsville, Virginia, and Bailey reads inbound the day it lands.

The platform

Hypha. The operator is the modeler.

An agentic AI platform that runs autonomous EDA, feature engineering, model selection, training, evaluation, and intel synthesis end to end. The operator describes the question. Hypha builds the model. The platform tells the truth about whether the model is ready to deploy.

Open a channel Read the assessment story

01 / 06

Agent fabric

Four named agents. One coordinated platform.

Hypha's architecture is the one PEO-SDA named in the published assessment. Four agents, one fabric, one operator in the seat. Each agent does one thing well. The fabric is what makes them coordinated.

Contextual Understanding Agent

Ingests, validates, and classifies operator data. The first agent the operator hands work to.
Model Building Agent

Runs autonomous EDA, feature engineering, training, and evaluation. Generates the model in near-real-time.
Semantic Fusion Engine

Merges real-time data streams with the operator's existing doctrine and reference material for predictive synthesis.
Visualization Agent

Composes the operator-facing dashboard with alerts, heatmaps, simulations, and the provenance the commander needs to make the call.

02 / 06

The model families

Pick the question. Hypha picks the model.

Hypha is task-agnostic. The pipeline is not one fixed model. It runs an autonomous EDA, preprocessing, feature engineering, model selection sweep that picks the right family for the data and the question the operator handed it.

Classification

"What kind of thing is this?" Activity recognition from sensor and track data, threat classification, vehicle and aircraft type ID, document triage, anomaly tagging.
Regression

"How much of something?" Time-to-target estimation, fuel and range prediction, casualty estimation, logistics demand forecasting.
Anomaly and novelty detection

"Is this weird?" Pattern-of-life violations, insider-threat indicators, equipment health, signal and comms intel.
Time-series and temporal models

Force buildup detection, attack-window forecasting, predictive maintenance on aircraft, ships, and vehicles.
Retrieval-augmented Q&A

Doctrine lookup with citations, OPORD and FRAGO drafting against the cited reference, lessons-learned mining across after-action reports, compliance Q&A.
Conversational onboarding and domain packs

Non-technical operators describe the problem in English. Hypha picks the pipeline.

03 / 06

The platform tells the truth

Honest evaluation, in the operator UI.

Most defense AI vendors do not ship a UI that says "this model failed, here is what to do." Hypha does. The canonical demo deliberately trains a classifier that scores 0.953 in cross-validation and 0.143 on a group-stratified holdout. The platform flags memorization, refuses to deploy, surfaces a DO NOT DEPLOY verdict, and auto-generates recovery guidance.

That is the credibility lever. In a community that rightly distrusts AI black boxes, the platform that says "this model is bad" is the platform that earns the call.

Cross-validation

0.953

Holdout (group-stratified)

0.143

Verdict

DO NOT DEPLOY

Recovery guidance

Auto-generated

04 / 06

The readiness ladder

TRL 6 today. TRL 9 by December.

The readiness ladder is what an evaluator wants to see, and it is visible. Hypha is at TRL 6 today, achieved April 2026 via USSOCOM TE 26-2. The target for sustained mission operations on a Department of War customer is December 2026.

TRL 1 / November 2025

Hypha concept formulated, AgentForge founded
TRL 2 / December 2025

Multi-agent platform architecture defined
TRL 3 / January 2026

End-to-end agentic POC demonstrated
TRL 4 / February 2026

Component validation in lab
TRL 5 / March 2026

Integrated platform validated in lab
TRL 6 / April 2026, achieved

USSOCOM TE 26-2 Agentic AI
TRL 7 / July 2026, projected

ARCYBER AI Effects at the Edge, Phase 1 prototype
TRL 8 / September 2026, projected

FedRAMP Moderate, ATO, CMMC 2.0 Level 2
TRL 9 / December 2026, projected

Sustained mission operations with a DoW customer

05 / 06

The workstation

One workstation. Air-gap capable. No cloud.

Hypha runs on a single 32 GB workstation. Reference deployment is 48 GB. Hardware-agnostic across Apple Silicon and ruggedized x86. The stack is local-first: Ollama with open-weight Gemma and Nomic, on-disk ChromaDB, in-memory BM25. No commercial cloud LLM. No API key. No network egress at runtime.

06 / 06

Scope

So no one buys the wrong thing.

Hypha is not a cloud LLM wrapper.
Hypha does not yet hold an ATO. Targeting H2 2026 via Second Front Systems' Game Warden inheritance.
Hypha does not require a forward-deployed engineer to stand up a model. That is the differentiator versus Palantir Gotham and Anduril Lattice.
Hypha runs on any 32 GB workstation. Retire any "Mac demo" framing if you have seen it elsewhere.

See the platform run on your data.

One inbox. One phone number. Bailey reads it himself.

Open a channel See the full architecture