Agent infrastructure
Infrastructure for software that acts on its own.
Agents hold credentials, call tools, and change production systems without a person watching each step. This is the ARC work for the part that still has to be checked: whether those actions can be trusted, verified, and governed.
This part of ARC
Readiness, a record, and a check.
The same company that runs the advisory work publishes these pages. Agent Readiness asks whether your systems can survive an agent. ARC Control, still with design partners, asks whether the intended change happened.
How should software that acts on its own be checked?
The system
- Measure
- Verify
- Govern
- Optimize
We check the system that was supposed to change.
Measure what happened. Verify it against the system that was supposed to change. Govern the authority it ran under. Optimize against cost per verified outcome rather than cost per token.
Agent operating lifecycle
- 01
Discover
Find out which capabilities, tools and resources exist, and which ones apply to the task in hand.
- 02
Authenticate
Prove which agent is acting, on whose behalf, and under which credential.
- 03
Understand
Read schemas, contracts and side effects well enough to call a tool correctly the first time.
- 04
Permission
Resolve what this identity is allowed to do right now, including the actions that need a human.
- 05
Execute
Run the work in an environment that fits its privacy, latency, policy and cost constraints.
- 06
Verify
Collect evidence from the systems that were supposed to change, not from the agent's own report.
- 07
Measure
Attribute cost, latency and success to a unit of work a business can recognise.
- 08
Remember
Persist state so the next run starts from what happened, not from an empty context.
- 09
Audit
Keep a record someone can defend later: what was attempted, under what authority, with what result.
The execution gap
Your agent said it finished. Did it?
A successful model response is not the same thing as a successful business outcome.
- An agent can complete a run while the payment failed.
- It can report that a record was updated when the destination system never changed.
- It can call the correct tool with the wrong authority.
- It can finish a workflow while violating the policy that should have governed it.
What the run reported
- Status
- completed
- Steps
- 7 / 7
- Tool calls
- 12 ok, 0 failed
- Latency
- 4.2s
What assurance asks
- What did the agent attempt?
- What authority did it use?
- What changed?
- What evidence exists?
- Was the outcome verified?
- What did the verified outcome cost?
Execution success is not a verified outcome.
Products
Four products, each with its maturity attached.
The runtime work, the protocol work and the individual scanners are deliberately not on this list. They sit underneath, in Developers and Labs, because presenting an experiment as a product is how trust gets spent.
ARC Control
Outcome assurance for production AI agents.
Understand what agents attempted, what authority they used, and whether the intended outcome occurred.
Where this is today: ARC Control is in design-partner development. It is not a product you can buy or sign up for yet. We are working through the verification model with a small number of teams running agents against production systems. There is no pricing page because there is nothing to price yet.
Agent Readiness
Make your systems usable by agents.
Evaluate whether agents can safely and reliably interact with your applications, APIs, tools and workflows.
Where this is today: Available in two forms. The scorecard is free and takes a few minutes. The full assessment is a two-to-three week engagement run by ARC against the systems you already operate, and it uses the open scanners in Labs as part of the evidence.
AI Fit
Determine where AI belongs before you build it.
Evaluate the problem, workload, economics, architecture and risks before committing engineering resources.
Where this is today: Available free on the web with no account and nothing stored. Deterministic rules match your description to a published catalogue of architecture shapes. A model never produces the recommendation or the score.
AI Fit Teams
Prepare the organization around the technology.
Understand AI readiness across teams, workflows, capabilities, governance and adoption.
Where this is today: Available as a three-to-four week assessment run by ARC. It is the organisational counterpart to AI Fit: the same scoring discipline, applied to the teams, workflows and habits that decide whether anything ships.
Flagship
ARC Control
Seeing a trace tells you what the agent attempted. Assurance asks a different question: did the intended outcome occur?
Where this is today: ARC Control is in design-partner development. It is not a product you can buy or sign up for yet. We are working through the verification model with a small number of teams running agents against production systems. There is no pricing page because there is nothing to price yet.
The operating contract
Agents need operating contracts.
An agent responsible for ongoing work needs more than instructions.
- Goal
- Trigger
- State
- Boundary
- Evidence
- Owner
A prompt is an instruction. A contract is something you can check afterwards and hand to someone else. The framework is free and published in full — there is no email gate on it.
- A defined objective
- What done looks like, in terms someone else could check.
- A trigger
- What starts the work, and what must be true before it does.
- Persistent state
- What it is allowed to remember between runs, and for how long.
- Enforced authority
- What it can reach, written as permissions rather than as guidance in a prompt.
- Evidence requirements
- What has to be produced to show the work was done.
- A human owner
- A named person. A team inbox does not count.
- Stop conditions
- What makes it halt rather than retry.
- Recovery behaviour
- What happens to work already in flight when it does.
Underneath
The layers that are not products yet.
Runtime, protocols, open tools and documentation. Each labelled with what it is, which in some cases is research.
Runtime
ResearchWhere a unit of AI work should execute, decided from capability, privacy, policy, latency and cost rather than from habit. Local, edge and cloud are a routing decision.
Protocols
LiveAgent-facing interfaces: MCP, agent-to-agent patterns, and the machine-to-machine conventions forming around them. We test against the published specs and publish what breaks.
Open tools
LiveScanners, catalogues and reference implementations from Labs. Free to run, no account, most of them store nothing.
Documentation
LiveArchitecture notes, schemas and implementation guides. Each tool documents itself on its own deployment.
Running agents you cannot yet verify?
Start with the free readiness scorecard, or book thirty minutes and we will tell you whether the problem is your interfaces, your authority model, or the fact that nothing confirms what changed.
A person reads every inquiry, usually replying within two business days.