Agentic systems
We test how agents plan, keep context, use tools, recover from errors, and coordinate over longer tasks.
All focus areas
Artificial Intelligence
Systems that can act, collaborate, and stay within clear boundaries.
We are looking closely at agentic systems: how they reason, use tools, work with other agents, and operate over time. The design challenge is making that work useful while keeping authority and limits visible to the organization.
AGENTA2ASECURITYCONTEXTENTERPRISE BOUNDARY
01 / The challenge
Useful autonomy needs a clear boundary.
An agent can plan a task, call tools, coordinate steps, and hand work to another system. Once it does that, authority, identity, data, and accountability become part of the design.
We turn those concerns into patterns an organization can use: systems with enough autonomy to help, clear limits on what they can do, and a record of the decisions they make.
02 / Current threads
The system around an agent.
We study the full system around an agent’s action: the agent, the protocols it uses, and the controls that keep its work within the right scope.
Protocols and coordination
We study protocols such as A2A and initiatives such as Agntcy, with a focus on discovery, communication, delegation, and clear responsibility.
Enterprise security
We work through identity, permissions, policy, audit trails, and safe tool access for agents inside an organization.
03 / Guardrails
An agent needs a defined scope.
The hard part is giving an agent enough room to help while keeping its authority clear.
Make authority explicit.
Tie the agent to a person, team, service, or approved workflow so its authority is clear.
Keep a record.
Keep decisions, tool calls, handoffs, and outcomes visible enough to review and learn from.
Set the limit.
Set clear limits for data, tools, scope, and escalation when conditions change.
04 / How we work
We take ideas into real systems.
01
Study the capabilities, protocols, and failure modes that shape agentic systems.
02
Test ideas against the practical realities of enterprise security, governance, and operations.
03
Build patterns into prototypes, guidance, and products people can use.
Related project
A governed route to enterprise AI.
AI Gateway puts these ideas into a product: a secure, multi-tenant route to model access for organizations. It is currently in closed beta.