Skip to content

Agent governance

Nothing sensitive moves without your say so.

Every Machine-like agent starts in draft mode. Anything that spends money, reaches a customer, or changes a record you care about stops at a gate until a person signs. Autonomy is earned one category at a time, and every action leaves a receipt.

Start free trial See what waits for you
Default
Draft mode on every new agent
Gated
Customer-visible writes and anything high risk
Silence
An unanswered request expires. It is never a yes.
Off
Pause any agent from your Terminal
Fig. 01 / Approval packet
Approval packet · AP-2214Illustrative · try it
  1. Drafted
  2. Dry run
  3. Our check
  4. Your call

R3Customer-visible write

Send Harbor Dental their proposal.

Action
Email the proposal, then update the deal
Systems
Your inbox and your CRM
Requested by
Sales follow-up agent
Exact change3 fields · 1 email
  1. deal.stageQualifiedProposal sent
  2. deal.next_stepnoneFollow up Thu, Oct 8
  3. email.tononedana@harbordental.example

WhyDana asked for pricing on Friday's call. The proposal uses your approved Q4 rate card and nothing else.

Call notes · Oct 2Rate card v4Deal record
Your decision
    Expires Oct 6, 5:00 PM. Nothing sends if it lapses.Waiting on you

    Founder-led teams
    run on Machine-like agents

    • fitDEGREE
    • Lake
    • espresso.ai
    • Customer Science Group
    • Course Careers
    • Alarm New England
    01GatesDefault policy for every new agent

    Every action has a risk tier.

    Before an agent touches anything, each action it can take is sorted into one of five tiers. The tier sets the default: run, land as a draft, wait for a person, or never happen.

    Gate tablePick a tool
    Inbox · default gates
    ActionTierDefault
    Read and sort new mailR0Runs
    Draft a reply in your voiceLands in your drafts folderR1Draft
    Label, file, or archive a threadR2Your policy
    Send a reply to a customer or leadR3Waits for you
    Email a list or send in bulkR4Blocked

    Starting pointCommon actions and their default tiers. Your exact table is set during scoping. Tighten any row whenever you like. Loosening one takes your signature.

    02AutonomyFour levels, set per category of work

    Autonomy gets earned.

    Nothing climbs without your signature.

    Levels apply to a category of work, not to the whole agent. An agent can send order updates on its own and still ask before it quotes a price. Every new agent starts in the bottom two.

    How a category climbs.

    Promotion is a proposal with evidence, never a setting someone flips. Revoking it takes one click, and the category drops back to draft mode on its next action.

    1. 01 / Evidence

      A clean record, attached.

      Machine-like proposes a promotion only for a category with a run of approvals you didn't need to edit and exceptions routed correctly.

    2. 02 / Limits

      Written limits.

      The proposal names the category and its ceilings: volume per day, who it may contact, records it may touch, and spend.

    3. 03 / Signature

      Your signature.

      Nothing changes until you sign in the Terminal. Each signature is logged with the policy version it created.

    Promotion · PR-0087Illustrative

    Order-status replies

    L1 Assisted to L3 Trusted

    Approved without edits38 in a row
    Exceptions routed correctly2 of 2
    Limit40 sends a day
    RecipientsExisting customers
    Spend$0
    03ReviewTwo keys on customer-visible work

    Two keys turn during review.

    Your policy can require two approvals on customer-visible and high-risk work: ours, then yours. By the time a packet reaches you, it has already passed a check.

    1. 01Agent

      Draft

      The agent prepares the work from your rules and sources.

    2. 02System

      Dry run

      A preview of the exact change, before and after. Nothing is touched yet.

    3. Key 1Us

      Our check

      Machine-like reviews the packet against your policy and sources.

    4. Key 2You

      Your approval

      Approve, request changes, or reject from your Terminal inbox.

    5. 05System

      Apply once

      The change is made a single time, even if a step retries.

    6. 06System

      Verify

      The agent confirms the change landed where it should.

    7. 07You

      Receipt

      The whole path, on the record.

    Where it lands
    Your Terminal inbox, with the full packet: the change, the reason, and the sources.
    What you can do
    Approve, request changes, or reject. Your note goes back to the agent with the packet.
    If you're away
    The request waits until it expires. Nothing runs on silence, and the expiry is logged.

    Who reviews first

    Agents you build yourself
    You are the first and only reviewer. You also find out what broke.
    Machine-like
    We check the work before it reaches you, and we own the fix when something goes wrong.
    04ReceiptsOne per run, kept for good

    Every action prints a receipt.

    What triggered it, what the agent did, who signed, and proof it landed. Written for a person, attached to the result it produced.

    Run 2214 · proposal_followup · Oct 5Illustrative
    1. ReadCall notes and deal record loaded1
    2. DraftProposal email drafted from rate card v42
    3. Dry runPreview built: 2 field changes, 1 email3
    4. CheckMachine-like review passed4
    5. GatedApproval requested from J. Ortiz4
    6. ApprovedJ. Ortiz approved with no edits4
    7. ApplySent once, deal updated5
    8. VerifyFound in sent folder, stage confirmed6
    Fictional names and recordsCompleted
    1. 1
      The trigger and the inputs

      What started the run and exactly what the agent read.

    2. 2
      Every step, in order

      Each tool the agent used and what it produced.

    3. 3
      The preview

      The exact change before it happened, with a before and after.

    4. 4
      The gate and the decision

      Which approvals were required, who gave them, and when.

    5. 5
      The write, once

      Applied a single time. A retried step can't send twice.

    6. 6
      The proof

      A check that the change is really there in your system.

    “Our partners at Machine-like gave us a more responsive support operation without forcing a massive migration. The agents worked across our systems and kept me in the loop when something needed my attention.”

    fitDEGREENick Hale
    CEO & Co-Founder, fitDEGREE    Read the fitDEGREE story
    05IncidentsLow risk is not zero risk

    If it runs off course, we'll get it back on track.

    The risk of an unintended action is low. It is never zero. What matters is what happens next and who handles it. That's Machine-like, not you.

    1. 01

      It stops.

      When a change can't be verified, the run halts and holds. The agent doesn't guess, and it doesn't retry its way into a second mistake.

    2. 02

      We hear first.

      Machine-like is alerted. You get a plain-language note in your Terminal explaining what happened and what's being done.

    3. 03

      It gets undone.

      Writes are built with a rollback path. A reversible change is reversed, and the reversal goes on the receipt.

    4. 04

      The rule gets fixed.

      We patch the workflow or the policy and test it again before the agent runs the next job. You never debug an agent.

    Incident · crm_stage_syncIllustrative
    1. Verify fail1 record didn't match its preview
    2. HaltRun stopped. 11 records untouched.
    3. AlertMachine-like notified
    4. RollbackRecord restored to its prior stage
    5. PatchField mapping fixed and re-tested
    6. Resume12 of 12 records verified
    Customer action required: noneRolled back

    Your off switch.

    Pause any agent from your Terminal. A paused agent refuses new work on its next action and stays paused until you turn it back on.

    Example agentAccepting work

    06ResponsibilityRules live in the system, not the prompt

    We manage. You decide.

    Your rules become policy the platform checks before every action. They are not instructions an agent tries to remember. An agent can't talk its way past a gate.

    You decideOwner

    The goals and the limits.

    • What done looks like for each job
    • Which systems the agent may reach
    • Limits on volume, recipients, records, and spend
    • The off-limits list: actions that never happen
    • Who approves what, and how fast
    • How much autonomy each category has
    We enforceMachine-like

    On every action, every time.

    • Checks each action against your policy before it runs
    • Blocks off-limits actions outright, with no override from the agent
    • Walls off your data from every other customer's, down to the database
    • Keeps your logins out of the agent's hands; you can revoke any connection
    • Logs every decision and every change to your policy
    • Owns the fix when something breaks
    Need the detail for a security review?See security policies
    07Draft your policyTwo minutes, no account

    Your rules are written down, never inferred.

    You get a one-page operating policy in plain English. It's yours to amend.

    • 01
      It's a real starting point.

      The draft maps to the same tiers and gates your agent runs under.

    • 02
      It shortens scoping.

      We start the call from your limits instead of a blank page.

    • 03
      It's yours to change.

      Tighten anything now. Loosen it later, with evidence and your signature.

    Operating policy · draftYours to keep
    POL-0001 · v1
    Tools it can reach
    How carefully to start
    Who approves
    Runs freely
    Reading and sorting mail, drafting replies; filing and labeling threads; looking up records, writing account briefs; adding notes and tasks.
    Waits for you
    Replies to customers or leads; deal stage or owner changes.
    Never
    Bulk or list email; deleting or merging records.
    Approver
    You approve every gated action. Unanswered requests expire.
    Autonomy
    Starts at Assisted. Moves up only by proposal, with evidence and your signature.
    Start free trial with this policy

    No credit card. Your policy is finalized on the scoping call.

    08QuestionsBefore you connect anything

    Asked before connecting.

    The questions owners ask before they give an agent access to anything.

    Q.01Will an agent ever email my customers without my approval?

    Not by default. Any message to a customer, prospect, or lead is a customer-visible write, and it waits for a person. A category of messages can send on its own only after Machine-like proposes a promotion with evidence and written limits, and you sign it. Bulk sends are blocked by default.

    Q.02If I have to approve things, what am I saving?

    The work itself. Reading, researching, and drafting never wait for you, so an approval is one decision on finished work, not a task. Approval is the starting point, not the steady state: as a category builds a clean record, we propose moving it up, and you approve less over time.

    Q.03What happens if I don't answer an approval request?

    Nothing. Every request carries an expiry. If it lapses, the action does not happen and the request is logged as expired. Silence is never treated as a yes.

    Q.04Can I turn an agent off?

    Yes. Pause any agent from your Terminal. A paused agent refuses new work on its next action and stays paused until you turn it back on. Revoking a promotion drops that category back to draft mode the same way.

    Q.05Who is responsible when an agent makes a mistake?

    Machine-like. The run stops when a change can't be verified, we're alerted, reversible changes are rolled back, and we fix and re-test the workflow before it runs again. You get a plain-language explanation in your Terminal. You never debug an agent.

    Q.06Can Machine-like or the agent see my passwords?

    You authorize each connection yourself, and your raw logins stay out of the agent's hands. You can revoke any connection at any time. Your data is kept separate from every other customer's at the database layer. Security policies have the detail.

    Q.07How does an agent earn more autonomy?

    One category at a time. Machine-like proposes a promotion only when a category has a clean record of approvals and correctly routed exceptions. The proposal names the limits: volume, recipients, records, and spend. Nothing changes until you sign, and you can revoke it in one click.

    Q.08Can I change the rules later?

    Yes. You can tighten any rule at any time. Loosening a rule takes your signature, and every change creates a new policy version on the record.

    Machine-like / Control

    Let's get back to the work of humans1

    Frontier labs build models. We put them to work, inside rules you set, so your team can get back to theirs.

    Start free trial See how our agents work