Intent provenance records where a machine-worker objective came from, who owns it, what evidence supports it, and how far it may travel.
Which human, policy, customer signal, experiment, dashboard, or upstream worker created the goal.
// AGENT OPS · TRUST DENSITY · June 20, 2026
Agent Ops Need Intent Provenance Before Workers Inherit Borrowed Goals
Autonomous work needs a memory of why the goal exists, not just a copied instruction that looks executable.
Autonomous systems pass work between workers faster than humans can inspect, and each handoff can strip away the reason behind the request.
Heads of Growth can let agents delegate work without letting borrowed goals become silent, stale, or commercially dumb.
The most dangerous instruction is the one nobody remembers approving.
Agent ops is becoming a chain of delegation. A scoring agent flags an account. A routing agent opens a task. A writing agent drafts the message. A RevOps agent updates the field. A customer agent changes the next step.
Each worker can be technically correct and still wrong. The original intent may have been narrow: test one segment, recover one stalled deal, suppress one sensitive cohort, or escalate only while a support issue is live.
If that intent does not travel with the work, the next agent inherits a command but not the judgment that made the command safe.
Intent provenance has five fields.
The commercial outcome the instruction exists to protect: revenue quality, customer trust, forecast truth, speed, risk reduction, or learning.
The source records, timestamps, confidence, customer facts, and assumptions supporting the action.
The segment, channel, time window, customer state, tool scope, and interpretation limits attached to the goal.
The human or control-plane role accountable for renewal, exception handling, contradiction, and shutdown.
Delegation without intent provenance turns agent teams into rumor mills.
A human rumor becomes weaker as it travels. An autonomous rumor becomes stronger because every receiving system treats it as structured work.
A note that began as "watch this renewal risk" becomes a suppression rule. A test offer becomes a field value. A one-week campaign becomes an outbound pattern. A single customer exception becomes a generalized playbook.
The problem is not that agents delegate. The problem is that the delegated work often loses its original why, then gains operational force as it moves downstream.
Goal ID
A durable identifier that follows the objective through every worker, tool, queue, and system-of-record update.
Intent statement
One plain sentence explaining what outcome the goal exists to create or protect.
Evidence packet
The records, citations, confidence, timestamps, owner notes, and known gaps that justify the instruction.
Translation rights
Which workers may reinterpret, narrow, expand, pause, or reject the goal during handoff.
Expiry rule
The date, event, performance boundary, or contradiction that forces the goal to retire or seek renewal.
Real example: the renewal-risk flag that became a growth tax.
A customer-success agent flags a set of accounts as renewal-risk because a product incident is unresolved. A GTM control plane correctly blocks aggressive expansion outreach for those accounts.
Two weeks later the incident is fixed. The risk label remains. An outbound agent inherits the old suppression rule, a forecast agent downgrades expansion probability, and a lifecycle agent delays executive follow-up. No single worker is failing. They are all obeying a borrowed goal whose intent expired.
With intent provenance, the suppression carries the incident ID, owner, evidence timestamp, affected accounts, renewal event, and expiry condition. When the incident closes, downstream workers know the goal must be reviewed instead of silently becoming policy.
A goal without provenance is not strategy. It is a rumor with permissions.
Count active agent objectives that have no human owner, original evidence, expiry rule, or current business reason.
Measure how often delegated goals change segment, channel, wording, or authority without an explicit translation receipt.
Track time between the business reason changing and downstream workers updating, pausing, or retiring the goal.
Install provenance before workers delegate work to each other.
Require every material agent task to carry source, owner, timestamp, and the upstream trigger.
Store the business reason in plain language so downstream workers can detect when literal execution violates intent.
Define whether each worker may execute only, narrow scope, reinterpret, expand, or stop the goal.
Make borrowed goals retire when the event, campaign, customer state, or owner assumption changes.
Concrete takeaway
Add intent provenance to one autonomous workflow where agents hand work to other agents.
- Name the origin: human, policy, customer signal, experiment, or upstream worker.
- Write the intent: the business outcome the goal exists to protect.
- Attach evidence: source records, timestamps, confidence, and missing context.
- Set translation rights: who can execute, modify, pause, escalate, or reject.
- Set expiry: the condition that makes the goal stale and forces review.
The blunt rule: do not let agents inherit goals unless they also inherit the reason, boundary, and owner behind them.