The worker treats old CRM fields, cached research, policy snippets, and operator notes as current until somebody catches the drift downstream.
// AGENT OPS · GOVERNED AGENT LABOR · June 6, 2026
Agent Ops Need Freshness Windows Before Machine Workers Act on Old Context
Old evidence is not evidence. It is a debt instrument with an expiry date.
The quiet failure mode is not hallucination. It is stale confidence.
Most teams worry that machine workers will invent facts. The more common operational problem is subtler: the agent uses a fact that was true, a rule that was approved, or a buyer signal that mattered yesterday and treats it as live authority today.
A GTM routing worker remembers that an account belonged to enterprise. A pricing assistant remembers that a discount was allowed. A research agent remembers that a prospect had a public hiring push. None of those memories are obviously wrong. That is why they are dangerous.
Freshness windows define how long context remains valid before a machine worker must re-check the source, ask for renewed authority, or downgrade itself into proposal mode.
The control plane sets different validity windows for policies, buyer data, approvals, pricing, risk signals, and source-of-truth writes.
Why one expiry date is not enough
Agent memory is not one substance. A company headcount scrape, a legal approval, a territory assignment, a campaign suppression list, and a human exception note all decay at different speeds.
Some context can safely live for a quarter. Brand positioning might change slowly. Other context decays by the hour. Buyer intent, outage status, deliverability health, budget approval, incident severity, and rep ownership can all flip before a machine worker finishes its queue.
The control-plane mistake is treating context as either trusted or untrusted. The better question is: trusted for what action, at what risk level, until what time?
Grounded example
The routing worker that acted on yesterday's account truth.
A revenue team gives an agent authority to enrich accounts and route high-fit leads. On Monday, the account record says a company belongs to mid-market, the buyer is expanding, and the assigned rep is available. The worker caches the profile because it has processed the same segment all week.
On Wednesday morning, the company enters an enterprise pilot, legal adds a restricted-contact note, and the assigned rep goes on leave. The agent still routes a warm inbound using Monday's context. Nobody notices until the wrong rep follows up with a message that violates the new account plan.
The failure was not a bad model. It was missing freshness policy. Account tier, contact restrictions, owner availability, and pilot status each needed a short validity window before any routing write or buyer-facing action.
Four freshness windows to install first
How long an approval, exception, or operating rule remains valid before the agent must reload the latest policy or ask for renewal.
How long scraped, enriched, CRM, product, support, or billing facts can drive action before the source of truth must be checked again.
How long buyer signals, usage spikes, hiring moves, funding events, and engagement activity remain strong enough to justify outreach or routing.
How long a temporary permission, exception budget, coverage handoff, or human instruction can authorize write, send, close, or escalation actions.
Freshness should change the action, not just raise an alert.
A stale context warning is weak if the agent can click through it. Freshness should be wired into the worker's allowed moves.
If buyer research is stale, the agent can draft but not send. If territory data is stale, it can propose routing but not assign ownership. If pricing approval is stale, it can prepare the deal note but not commit the discount. If a suppression list is stale, it must stop buyer-facing work entirely.
This is where governed agent labor becomes more than observability. The system does not merely display risk. It converts expiry into reduced authority.
The freshness matrix
Context class
Example
Default action
Buyer signal
Intent spike, funding news, hiring plan, product usage surge, webinar attendance.
Refresh before outreach.
Operating policy
Discount rule, compliance note, escalation threshold, suppression policy, routing exception.
Reload before write.
Ownership
Rep assignment, customer-success owner, territory, named account status, partner conflict.
Verify before routing.
Temporary authority
One-off approval, coverage handoff, incident permission, exception budget.
Expire aggressively.
Machine workers should not be allowed to turn old context into new authority.
Where freshness pays first
Three lanes with fast context decay
Routing and ownership
Territory, account tier, rep capacity, and conflict rules decay quickly. Require fresh source-of-truth reads before assignment writes.
Outbound timing
Intent signals expire. A machine worker should not send confident outreach from old funding news, stale hiring data, or last week's engagement.
Temporary permissions
Coverage, exception, and incident permissions should expire by time, queue, or receipt count instead of becoming ambient power.
Concrete takeaway for this week
Pick one autonomous worker that uses remembered context to make or propose decisions. Add freshness windows to the context that can damage trust when stale.
- List context classes: policy, data, intent, ownership, authority, and human notes.
- Set expiry by risk: buyer-facing, revenue-impacting, and source-of-truth writes get the shortest windows.
- Bind expiry to permissions: stale context should force draft, proposal, refresh, escalation, or stop modes.
- Log refresh receipts: record source, timestamp, policy version, and action allowed after refresh.
- Review exceptions: if operators keep extending a window, the workflow needs a better source-of-truth feed.
The blunt rule: if context can authorize machine work, it needs an expiry date before the agent gets to act.