Architecture
Protocols and the core compiler.
Two protocols carry every task: A2A connects the agents, the ATO carries the work. Both are audited before anything is executed. The core compiler reads your core signature and compiles the system into repo-ready code on any supported provider.
A2A
Agent to agent, inside and out.
The same connection serves application agents, CLI agents, teams and external entities.
- In-app agents.Application agents hand each other work over A2A within a core: intake to analysis, analysis to reconciliation, with the task's audit attached.
- CLI tiers.Development and operations tiers exchange tasks over A2A; a tier-1 request becomes a tier-2 task that waits for approval.
- Teams.Cores of different teams connect through the console, never directly; membership and role decide what crosses.
- External entities.Partners and client systems connect as named external agents with their own identity, scope and channel.
- Private tunnels.Connections are outbound-only from the core over a private, mutually authenticated tunnel. Nothing listens on the client side.
- Protection.Every connection carries an identity, a scope and an allowlist of routes; unknown agents, expired grants and out-of-scope calls are refused and recorded.
- Custom data and connection protocols.Exclusive data formats and connection protocols compile as custom tools on the same graph, under the same identity, scope and audit.
ATO
The task carries its own audit.
One agent task object per task: the data path, the execution order, the tokens spent, the alerts raised. Every byte that matters is audited and processed by logic before any model sees it.
- Data path.The ATO records where each input came from and where each output went; the path is part of the object, not reconstructed.
- Execution order.Steps run in the order the ATO states; a step cannot start before its predecessor is recorded.
- Tokens and cost.Model usage per step is written to the ATO and rolled up per agent and per core in æ-c os.
- Alerts.Approvals, refusals and threshold breaches are ATO events with a named recipient; nothing waits silently.
- Logic first.Deterministic checks run on every ATO before a model is called: schema, policy, budget, scope.
- Rhythm drift.A customizable ML model watches the task rhythm per agent: timing, volume, retry and cost patterns. Drift raises an ATO alert; the model and its thresholds are the client's to tune.
- Workers and sessions.Each ATO runs in one session of one worker; both are recorded with the task.
- Isolation.Each worker runs in its own sandbox with allowlisted routes; nothing else is reachable.
Custom protocols. Exclusive protocols and complex algorithms are welcome as custom tools: they compile into the same graph, ride the same ATO and write to the same audit channel.
Where to go next
Read the published shape of the core record the compiler reads, the vocabulary these protocols are written in, or the delivery chapter that binds a repository to its framework release.