Brane

Agent Mesh thesis

agent mesh: a unified control plane for ai agents

This is the long-term thesis. For what Brane does today, see the product page →

Agents do not only call services. They choose models, invoke tools, read memory, write state, delegate work, run code, and create side effects. Brane starts with model and capability decisions, then grows into the governed action layer underneath agent frameworks.

Positioning

LangGraph, CrewAI, OpenAI Agents SDK, MCP servers, A2A agents, and internal tools help agents act. Brane gives those actions a common programmable loop: capability, policy, decision, trace.

The capability policy is the enforcement point. It runs before the action executes. The same policy that governs a tool call today can govern retrieval systems, workers, sandboxes, memory stores, APIs, humans, and other agents.

The action loop

01observe

workflow, step, actor, requested capability, data class

02decide

model, tool, retrieval route, sandbox, human approval

03enforce

local SDK, gateway, worker boundary, policy bundle

04trace

decision, invocation, fallback, cost, latency, side effect

05learn

outcomes update routes, policies, caches, and evaluations

Expansion path

Today, Brane wraps agent steps as capabilities, runs policy functions before they execute, and enforces allow or deny decisions with full traces.

Next, the same policy loop extends to knowledge bases, sandboxes, model selection, tool routing, and human approval queues. Every resource an agent can reach exposes a governed capability. Policies decide what agents can use, under what constraints, with full traces.

python
# Tomorrow: same policy primitive, applied to retrieval

@brane.policy("knowledge.query")
def restrict_sensitive_collections(ctx):
    if ctx.collection.contains_pii and not ctx.session.is_authorized:
        return Decision.block(reason="unauthorized PII access")
    return Decision.allow()

The private beta starts with capability policies and decision traces, then expands the loop one component at a time.