Glove Foundry
Foundry is to agent systems what an application framework is to the web: conventions, file routing, composition, a runtime, a development server, typed boundaries, and an inspection workbench in one coherent stack.
Foundry does not replace Glove. Glove provides the agent loop and capabilities. Foundry turns those capabilities into applications made of definitions, persisted instances, conversations, event-driven playbooks, schedules, and observable runs.
The mental model
Definition
The typed set of things an agent may become. Identity comes from the file route.
Instance
A persisted agent identity with installed apps, playbooks, schedules, configuration, and context.
Conversation
A durable stream of messages and work. An instance can participate in many conversations.
Run
One context-aware assembly and execution, correlated across passes, tools, events, and artifacts.
Purpose-built boundaries
| Foundry owns | You provide |
|---|---|
| Definition discovery, file identities, assembly, run lifecycle, inspection | Model provider configuration and API keys |
| Instance, conversation, playbook, schedule, and installation contracts | Durable adapters for the persistence guarantees you need |
| Application and transmission surfaces exposed to agents | Credential acquisition, selection policy, storage, and refresh |
| A stable agent-system vocabulary over backend execution | Transport, identity, authorization, and deployment policy |
Backend infrastructure details stay behind adapters. Foundry talks about inbound transmissions, workers, schedules, and connections—not the implementation vocabulary of any particular execution backend.
The workflow is allowed to emerge
Foundry supports deterministic handlers and predefined schedules, but it does not require the entire job to be expressed as a static graph. An agent can interpret a request, split it into campaigns, provision peers, open shared work, create artifacts, schedule a follow-up, sleep, and reconvene. The boundaries are typed; the route through them can respond to intent.
Read by job
Definitions & instances
What belongs in code, what belongs in data, and how persisted instances are reconstructed.
Read →Agent composition
Message-aware lazy assembly, shared tools, models, memory, inboxes, layers, MCP, and mesh calls.
Read →Apps & transmissions
Dynamically installed capabilities, inbound and outbound routes, accounts, and credential adapters.
Read →Playbooks & schedules
Runtime playbooks, event subscriptions, instance provisioning, future triggers, sleep, wake, and cancellation.
Read →Conversations & work
Multiple conversations per instance, inboxes, shared workspaces, tasks, artifacts, and background work.
Read →Working environments
Mount a VFS, REPL, skills, and document or media adapters for artifact-producing agents.
Read →Multi-agent systems
Layered agents, subagents, S2S and S2V calls, fan-out, handoffs, and shared work.
Read →Runtime & HTTP API
Typed clients, handlers, activation endpoints, Effect services, and adapter boundaries.
Read →Inspector & observability
Follow runs, model passes, tool calls, messages, retries, artifacts, and safe working notes.
Read →Deploy
Production adapters, durable state, Foundry workers, and the path from local Foundry to cloud agents.
Read →