Multi-agent harness and tenant API
AvenFort builds custom agents for company clients. The agents differ; the way they are allowed to touch the product should not. I own the Python harness and the .NET API boundary those agents run through — tools, tenancy, and who may call what.
One harness, many client agents
The company is young. There is no legacy platform to migrate. What exists is a pipeline of client work: one team shipping agents, another trying to show those agents to companies that might buy. The agents are not one product. They are per-client systems — merger research that compares companies, satellite review of cacao plots, work for law firms, and whatever the next contract needs.
If each of those lands as its own stack, you get a folder of demos and no way to say which user may run which agent, or which tools that agent may call. The product question is not “can we prompt an LLM.” It is whether the second and third client can attach without copying the first agent’s accidents.
- Several developers build independent client agents in parallel. Speed is rewarded. Shared rules are not, unless they live in the harness.
- The loop is Python. The tenant API is .NET — the language the team already wanted for product work. Agents call that API; they do not bypass it.
- Runtime is an Azure VM. Secrets, process, and disk stay there. They should not leak into tool names.
- Work is still being shown to clients. There is nothing to inflate: no traffic number, no closed-won count. The page is the architecture, not a launch.
Three surfaces
agents → python harness
↓
.net tenant api
↓
azure
Agents are policies. They decide what to try next for a given client. They should not know how a tenant stores a record or how auth is enforced.
The harness is the loop: session, tool dispatch, retries, and refusal. It is also where access is decided — which users may use which agents, and which tools those agents are allowed to call. If that lives in each agent repo, tenancy becomes folklore.
The tenant API is the product. Validation, persistence, and domain rules stay in .NET. The harness is a client of that API. An agent that can skip a rule a human user cannot skip is not a tool.
Other engineers write the client agents. My job is the quality of the path those agents run on: reviews, housekeeping in the harness, and the contracts for tools and access. When the rest of the team is shipping to show a client something this week, the harness is what keeps that week from becoming the platform.
I also work on the .NET API side of that boundary. Two languages is not a talking point. It is the split the company chose; the work is to keep it honest.
Python harness · .NET tenant API · Azure VM · per-client agents (M&A comparison, cacao satellite review, legal, and other custom work)
Client agents can be added without forking a new product. Access and tools are enforced in one place. The company can take those agents to conversations with companies without each demo inventing a new way to mutate tenant state. Numbers are not the story yet; the boundary is.
- If developers can ship an agent that talks to the database, they will. Put tool access in the harness before the third client arrives.
- A polyglot split only works if the API is the product and the loop is a client. Convenience helpers that accept natural language and choose a table will not survive a second agent.
- Housekeeping is the job. Reviews, tenancy, and “you may not call that tool” are how a small team stays a platform instead of a pile of demos.
Reviewing or building an agent platform?
I care about harness quality, tool access, and tenancy more than prompt theater. Open to remote technical-lead conversations.