The agent is named before it acts
Trust starts when the user gives the agent an identity. There is no generic default assistant pretending to own the relationship.
White paper
Agent First: software built around conversational execution
The next software shift is not a chat panel attached to an app. It is a product architecture where a named agent becomes the operating presence for the user, carrying memory, policy, tools, and execution across every surface where work happens.
Thesis
Chatbots wait. Agent First systems execute.
Most AI products still treat the model as a feature. A user types into a box, gets a response, and then leaves the conversation to do the real work somewhere else. That pattern cannot become the operating layer for a business.
Agent First software starts from a different premise: the user should be able to speak to a named agent and have that agent carry the task through memory, tools, policy, execution, verification, and approval. The interface is conversational, but the product is operational.
Clara Agents is built on this premise. Rian is not a chat widget that sends requests to a hidden automation platform. Rian is the Clara Harness presence across the website, backend, shared EC2, Cloudflare Workers, build farm, preview path, PR path, deployment path, and the wider Heru ecosystem.
Industry shift
The Responsive Web Design moment for AI software
Responsive Web Design changed how teams thought about screens. The breakthrough was not one layout trick. It was a new default: software had to adapt across the surfaces people actually used.
Agent First applies that kind of shift to execution. AI software has to adapt across the surfaces where work actually happens. The agent cannot live in one chat box while the real product, code, content, approvals, and deployments live somewhere else.
Harness core
Frameworks become surfaces around the agent
Agent First does not throw away Next.js, React Native, Express, Cloudflare Workers, or the other frameworks teams already use. It changes what sits at the center. The Agent Harness becomes the core control plane for identity, memory, tools, policy, execution, audit, and approval.
The frameworks still matter, but they become surfaces around the agent. The website renders the experience, the mobile app carries the voice session, the backend enforces contracts, Workers handle edge execution, and the build farm proves code changes. The named agent remains the continuous operational presence across all of them.
Open blueprint
Agent Boilerplate makes the pattern public
The next step is Agent Boilerplate: a refactored, public, open-source starting point distilled from lessons learned building real Agent First systems. It should teach engineers how to build from the Harness core out instead of adding AI as a feature at the edge.
The public blueprint should remove private product IP, customer context, operational credentials, deployment specifics, and Clara-only implementation details. What remains is the portable pattern: Harness-core architecture, named agents, memory, tools, policy, cross-surface execution, audit trails, approval gates, and adapter examples for harnesses such as PI, Hermes, and Clara Code.
As more harnesses become open and interoperable, the boilerplate should make room for them. The point is not to lock developers to one implementation. The point is to make Agent First the default mental model for AI infrastructure.
Principles
The architecture has to change
Memory belongs to the product
The agent reasons through a shared per-Heru brain so context survives sessions, surfaces, and tools without becoming a loose chat transcript.
Execution crosses every surface
The same agent presence works through the website, backend, shared EC2, Cloudflare Workers, build farm, PRs, previews, deploys, and other Herus.
Every serious action leaves evidence
Content publishes, engineering changes, deployments, and approvals produce traceable artifacts instead of vague claims that work was done.
Cross-surface execution
Ubiquity means the agent is the runtime presence
The cross-surface execution layer is the core moat. It is what lets the same named agent operate where the work actually lives: content systems, user dashboards, backend services, worker environments, code repositories, build farms, preview deployments, and approval gates.
Execution loop
- 1The user speaks to their named agent in the product.
- 2The agent understands the request through memory, policy, and tool context.
- 3The same agent presence executes across the required surface.
- 4Work produces concrete artifacts: draft, preview, branch, check, PR, or deploy result.
- 5The owner approves the final state before it becomes live.
Content work
A business owner should be able to ask for a contact page and get a real draft, preview URL, editable sections, owner approval, and a public page. The agent should remember what was changed and keep draft edits isolated until publish.
Engineering work
An engineering request is not complete when an intent is stored. The agent has to create real work: branch, worker assignment, code change, checks, preview URL, PR, approval trail, deployment result, and a trace written back to the brain.
Glossary
Terms for the Agent First era
Agent First
An application architecture where conversational AI agents are primary actors that understand, act, verify, and report evidence across the product.
Cross-surface execution layer
The runtime layer that lets a named agent operate across frontend, backend, workers, build systems, repositories, previews, deploys, and other Herus.
Clara Harness
The core agent runtime. Rian does not invoke Clara Harness as an outside system; Rian is Clara Harness embodied across the surfaces she operates on.
Harness-core architecture
A product structure where Next.js, React Native, Express, Workers, and similar frameworks remain important surfaces, but wrap around the Agent Harness control plane.
Agent Boilerplate
The public, open-source blueprint for Agent First applications, refactored from lessons learned and generalized around pluggable Agent Harness adapters.
Traceable dispatch
Engineering execution that creates real proof: branch, worker, logs, checks, preview URL, PR, approval record, deployment result, and brain trace.
The standard is proof, not promises
Agent First platforms will be judged by whether their agents can carry work across surfaces and come back with evidence. Clara is building that standard into the product from the foundation.