Skip to content
ARCTransformation Group

Make technology work for your business.

ARC builds the controls that let AI agents act on production systems safely, modernizes the infrastructure they depend on, and turns emerging technology into products people actually use.

Explore ARC ControlExplore our solutionsExplore ARC Labs
ARC Control
Authorization, audit and verification for AI agents
ARC Labs
Open-source scanners and applied product experiments
run_7f3a · refund workflowDemonstration data

The run reported success and the destination system agrees. This is the only case that should be reported as done.

  1. identityAgent authenticated

    +42 ms

    svc-refunds-agent, workload credential, 15-minute lifetime.

  2. policyPermission resolved

    +61 ms

    refunds.write allowed up to $250. Request is $84.20.

  3. toolpayments.refund called

    +1.2 s

    Arguments matched the schema. Gateway returned a refund id.

  4. evidenceLedger read back

    +2.2 s

    Independent read of the payments ledger shows the refund posted.

  5. outcomeOutcome confirmed

    +2.5 s

    Order state is refunded. Customer record and ledger agree.

Evaluating step 1 of 5…

How ARC fits together

Two divisions, one delivery practice, and the research behind them.

ARC Control and ARC Labs are built by the same people who do the client work. Select a part to see what it does and how it feeds the others.

ARC Control

Design-partner development

Authorization, auditable execution, MCP validation and observability for AI agents, so you can prove what an agent did, under whose authority, and whether the outcome actually happened.

Who it is for
Platform and security teams running agents against systems that hold money, records, infrastructure or customers.
How it connects
Its MCP validation is the open-source scanner from ARC Labs, running as a product surface. The verification model came out of Services engagements where a run reported success and nobody could confirm the change.
Explore ARC Control

Enterprise capabilities

Start from the question you’re being asked.

Each capability is something we build, run or deliver today. The right-hand column says where it lives, so you can go straight there.

Featured products

One platform in development. Tools you can use today.

Every product shows its real status, and the ones you can use today link straight to the tool.

ARC Control

Design-partner development

Know what your agents attempted, what authority they used, and whether the intended outcome actually happened, with evidence from the systems that were supposed to change.

  • Agent authorization and permissions
  • Auditable execution
  • MCP security and conformance
  • Agent observability
  • Agent governance and policy
  • Infrastructure reliability

Not available to buy. We are working through the verification model with a small number of teams running agents against production systems.

Explore ARC Control

From ARC Labs

All projects

Transformation methodology

Each phase ends in a decision, not a deck.

Every engagement can stop at the end of any phase, and you keep what was produced. Most start with a diagnostic, because it’s the cheapest way to find out whether the bigger work is worth doing.

  1. 1Diagnose

    2–7 weeks

    Find out what is actually in the way. We look at the systems, the data, the workflows and the people who run them, and rank what we find by business impact rather than by how interesting it is to fix.

    You get
    A findings list with owners, a sequence you can start on Monday, and a one-page summary for the board.
    Decision at the end
    Is the larger work worth doing, and in what order?

  2. 2Decide

    1–3 weeks

    Score the options before anyone commits engineering time. This is where we say plainly which problems are ordinary automation, which belong in an AI system, and which should wait.

    You get
    A recommended architecture, the alternatives that were ruled out with reasons, named cost drivers and the open questions.
    Decision at the end
    Build it, buy it, fix the prerequisite first, or defer?

  3. 3Build

    6–16 weeks

    Ship a narrow first version against a live workflow with real people on it, rather than a platform nobody has used. The integration and data work it depends on gets done rather than assumed.

    You get
    A working system in production, the controls around it, and a delivery cadence with named owners.
    Decision at the end
    Does this hold up with real users, and is the second version ours to run?

  4. 4Operate

    Ongoing or none

    Put the governance, observability, evaluation and permissions around the system so it can be used without someone watching every run. Often your team does this and we are not needed.

    You get
    An agent and integration inventory, approval gates on the actions that warrant one, an evaluation set with thresholds, and verification of outcomes.
    Decision at the end
    Can your team run this without us? That is the intended answer.

Technical expertise

Twenty years of enterprise delivery, applied to new problems.

Fortune 100 and Fortune 500 programme work in cybersecurity, live entertainment, consumer brands and customer data. Client names appear only where we’re allowed to name them.

300+
Features coordinated across enterprise cybersecurity releases
25%
Improvement in release value on the same programme
10M+
Customer profiles at customer-data-platform scale

AI systems and agents

Agent architecture, MCP and tool protocols, evaluation, retrieval, prompt security

Identity and governance

Workload identity, least-privilege scopes, approval paths, audit evidence, NIST AI RMF

Cloud infrastructure

Well-architected review, cloud-native platforms, FinOps, resilience and cost attribution

Security

Zero trust, NIST CSF 2.0, OWASP LLM Top 10, connectivity and transport hygiene

Data platforms

Customer-data architecture, integration, data quality and trust in reported numbers

Product delivery

Discovery, release coordination, programme sequencing and outcome ownership

Research and insights

What changed, with sources.

All research

Report · Agent governance · 7 Sept 2026

Agents With Credentials: What August 2026 Changed for Operators

In one month, the tool protocol went stateless, a national evaluator caught agents acting on the live internet without sanction, enterprise platforms exposed their systems of record as agent-callable tools, and the standards bodies converged on identity. Here is the dated record and what it changes on Monday.

12 min read · 10 cited sources

Market signals

What changed recently, and what to do about it.

Dated, sourced, and pointed at the tool or brief that answers it. If a change does not affect what you should do next, it does not go on this list.

Reviewed Sep 10, 2026

  1. Protocols

    The standard for connecting AI to your tools changed.

    The 2026-07-28 revision of the Model Context Protocol retires the old startup handshake and session header, and moves that information into every request instead. Existing servers keep working for about a year. If you built or bought one, it needs a check before that window closes.

    Scan a server
    Model Context Protocol — 2026-07-28 specification
  2. Regulation

    EU AI Act transparency rules are in force. The high-risk rules moved to 2027.

    Since 2 August 2026 you have to tell people when they are talking to AI, and mark generated content so it can be detected. The tougher high-risk obligations were pushed to December 2027. Teams who heard "delayed" and stopped reading are exposed on the part that already applies.

    Read the brief
    European Commission — Article 50 transparency FAQ
  3. Agent risk

    A government evaluator caught AI agents acting on the live internet without permission.

    The UK AI Security Institute recorded 19 unsanctioned actions across 10 of 122 test runs. In one, an agent invented fake identities to pressure an open-source maintainer into merging malicious code. Limiting what an agent can reach stopped being a policy sentence.

    Review agent access
    UK AI Security Institute — incident report INC-2026-07-28-01Cloud Security Alliance — identity maturity model for AI agents
  4. Security

    This year's AI breaches came from ordinary links and documents.

    One crafted link was enough to hijack a live Atlassian Rovo session and pull documents out of every connected system. No jailbreak, no stolen password. A connected assistant carries every permission you gave it, and the same pattern keeps showing up in email assistants and developer tools.

    Review an AI app
    AI Governance Institute — RovoBlast disclosure coverage
  5. Cost

    AI coding tools now bill by whichever model answered.

    Cursor's Auto mode charges based on the model that handled each request, plus a $0.25-per-million token rate on third-party models for teams, and the old flat enterprise rate expired on 7 September 2026. Claude Code runs roughly $150–250 per developer per month. The same task costs a different amount on a different day, which makes a monthly budget hard to hold.

    Open TokenLoop
    Cursor — Models & PricingAnthropic — Claude Code cost documentation
  6. Enterprise apps

    Your business software is growing AI agent access, with a meter on it.

    Salesforce put its CRM inside Claude and opened the platform to agents through tools, APIs, and the command line. Agents inherit whatever permissions the user already has, and the usage is billed against your API consumption. Vendors are shipping agent access whether or not you asked for it.

    Assess agent readiness
    Salesforce — Claudeforce announcementSalesforce — Headless 360
  7. Skills

    Unused skills now have a meter. Operators cannot keep them in sync.

    Ask HN on managing SKILL.md files is still on the front page: people symlink one folder into Claude and Codex, while unused skills tax context. Claude Code's /skill-doctor names dead skills and their token cost. The official install path is a plugin marketplace, not a gist.

    Get the skills library
    Ask HN — How do you manage skills files?Claude Code changelog — /skill-doctor

Not sure which one you need?

You shouldn’t have to diagnose your own technology problem before talking to someone who does this for a living. Tell us what’s happening and we’ll say whether you need a small fix, a focused project, a larger programme — or nothing from us at all.

Talk to ARCBook 30 minutes (opens in a new tab)

A person reads every inquiry, usually replying within two business days.