Corpus Agentis
The field book to agent ecosystems
The field book to agent ecosystems
Catalogue · Modalities

The four modalities of deployment

Where an agent runs determines who controls its data, cost and reach. The Fieldbook tracks four deployment modalities, local, cloud, caged and headless, to show how the environment around an agent shapes its practical capabilities and constraints.

LocalCloudCagedHeadless
Deployment modality

Local

The agent runs entirely on the user's own hardware: full data control, no egress, capped by local compute.

Examples
llama.cpp · Ollama · LM Studio · OpenHands + Ollama
Limitation
Capability ceiling set by local silicon; no elastic scale
Deployment modality

Cloud

The agent runs on managed infrastructure: serverless, sandboxed containers, or a fully managed platform.

Examples
Cloudflare Workers AI (serverless) · AutoGen/OpenHands Docker (sandboxed) · Bedrock AgentCore · Agentforce (managed)
Limitation
Data leaves the perimeter; egress + token cost; platform dependency
Deployment modality

Caged

An agent locked inside one vendor's product surface: usable only through that vendor's UI, not callable as an external agent. The platform lock-in problem.

Examples
Early HubSpot Breeze · Salesforce Einstein · early Agentforce
Limitation
No external API/agent access; you rent the agent, you can't orchestrate it
Deployment modality

Headless

Agent-first: the product is exposed as an API with no required UI: "the API is the UI." Other agents can drive it directly.

Examples
Salesforce Agentforce Headless 360 (Benioff, Apr 2026: "No browser required. Our API is the UI.")
Limitation
Emerging; most SaaS headless roadmaps undated
Field notes

Where an agent runs sets who controls its data, cost and reach

Deployment is an architectural decision, not an implementation detail. Local agents run on an individual's device, keeping the person close to the work and its controls. Cloud agents run as managed services, with providers operating the infrastructure and inference. Caged or vendor-locked agents are contained within a vendor's product surface. Headless agents run without a browser, operating through services, tools and APIs.

Show more

Autonomy changes the deployment model. Today 96 of 103 catalogued agents are cloud-deployed and only one is headless, which shows how early that shift still is. Manual agents are productive but they pause when the operator does: run under close supervision from a laptop and the work stops when the person steps away. Cloud and headless architectures make longer-running, less supervised execution possible. A useful rule of thumb is that supervised agents run locally or within a platform, while unsupervised work moves to the cloud or goes headless. The transition should be intentional, because greater autonomy needs stronger trust, clearer permissions and governance that matches the consequences of the work.

From the corpus, curated by Brandon Chaplin
Common questions
What is a headless AI agent?

An agent with no interface of its own, driven by other software through an API instead of a person clicking a screen. It is how agents get called by a backend service, a scheduled job or another agent.

Where do AI agents run?

Four places, in practice: on a local machine, as a cloud service, inside a vendor’s product, or headless behind an API. Almost everything in production today is cloud-hosted, because the work needs to carry on when nobody is watching.

Should an AI agent run locally or in the cloud?

Locally when the data must not leave the machine or you want to supervise each step. In the cloud when the task is long-running or has to continue unattended. A local agent stops when the person does, which is the limit most teams hit first.

Can I run an AI agent on my own computer?

Yes. Agent frameworks run fine on a laptop and can call either a hosted model API or a local open-weights model. Running the model locally as well needs a lot of memory and a capable GPU, and it will be slower than a hosted equivalent.

What is vendor lock-in with AI agents?

When an agent only runs inside one company’s product, on that company’s models, data and pricing. It is convenient and awkward to leave, because the prompts, tools and integrations rarely move to another vendor without being rewritten.

Next in the learning path