Durable agentic operations

Agents that stay on the jobuntil the outcome is done.

Run agentic workflows for hours, days, or months. Pause safely between events, execute inside isolated environments, recover from failure, and bring humans in only where judgment belongs.

Durable state Isolated execution Human approvals Outcome tracking

Active workstream

return-to-sender / YSE-1048

Waiting safely
01

Carrier signal received

Completed

02

Order and customer matched

Completed

03

Resolution plan generated

Approved

04

Customer response

Waiting · 18h

05

Execute and verify outcome

Queued

Elapsed

2d 7h

Active compute

19m

Budget used

31%

Hours → months

Workflow lifetime

Pause → resume

Without losing state

One → many

Agents in parallel

Every step

Evidence and audit

One operating model

Durable orchestration meets isolated execution.

StateSet separates the lifecycle of the outcome from the temporary compute used to do the work. The result is adaptive reasoning inside deterministic operational boundaries.

Remember what must happen

iCommerce Workstream Engine

A Rust-based orchestration layer built on Temporal owns state, timers, retries, signals, child workflows, approvals, and completion.

  • Durable state and recovery
  • Deterministic policy gates
  • Parallel activities and task queues

Give agents a safe computer

iCommerce Sandbox Engine

Each task receives an isolated environment for code, files, browsers, approved services, scoped secrets, and generated artifacts.

  • Kubernetes-native isolation
  • Checkpoints and session rotation
  • Cost, iteration, and compute budgets

Keep operators in control

StateSet Control Plane

Monitor active work, enforce authority, intervene at any time, and measure the business result instead of merely logging activity.

  • Pause, steer, approve, or cancel
  • Evidence and audit timelines
  • Outcome and budget tracking
01

Signal

An event or objective starts the workstream.

02

Execute

Specialized agents work in isolated sandboxes.

03

Wait

Timers and external events pause work safely.

04

Resume

Retries, signals, or decisions move work forward.

05

Outcome

Actions, evidence, and business results persist.

High-value workstreams

Start where waiting, coordination, and recovery matter.

The best first workflow is not the one with the most prompts. It is the one where durable ownership removes operational delay and creates a measurable result.

01

Commerce operations

Return-to-sender resolution

Own an RTS exception from the first carrier signal through customer recovery, refund, reshipment, or outreach.

Measured outcome

Recovered revenue, resolution time, and final disposition

What the agent owns

  1. 1Detect returned shipments and reconstruct carrier events.
  2. 2Match the order, customer, inventory, and applicable policy.
  3. 3Wait for carrier or customer responses without losing state.
  4. 4Escalate ambiguity and track recovered revenue through completion.
02

Customer experience

Customer-support resolution

Investigate and resolve complex cases across the helpdesk, commerce platform, email, warehouse, and carrier systems.

Measured outcome

Verified resolution, CSAT, retained revenue, and cost per case

What the agent owns

  1. 1Gather ticket, order, payment, and shipment context.
  2. 2Run reconciliation or diagnostic scripts in an isolated sandbox.
  3. 3Request missing information and resume when the customer replies.
  4. 4Execute an approved resolution and verify the outcome.
03

Finance operations

Chargeback and dispute management

Build, approve, submit, and monitor an evidence-backed dispute response over its complete lifecycle.

Measured outcome

Win rate, recovered revenue, response time, and evidence quality

What the agent owns

  1. 1Detect the dispute and calculate the response deadline.
  2. 2Collect delivery, order, identity, and communication evidence.
  3. 3Assemble the submission package and route material cases for approval.
  4. 4Monitor the dispute for weeks and record recovered revenue.
More durable workflows

One platform. Many outcomes that outlive a session.

Use the same operating model anywhere work must investigate, act, wait, recover, and prove completion.

Days to months

Accounts receivable

Reconcile accounts, run controlled follow-ups, propose payment arrangements, and track collected revenue.

Hours to days

Production incident investigator

Test hypotheses in isolation, execute approved remediation, monitor recovery, and produce the incident record.

Hours to weeks

Software maintenance

Implement and test changes, open a pull request, wait for CI and review, then address feedback until merge-ready.

Days to weeks

Security remediation

Reproduce findings, build patches, coordinate approvals, monitor deployment, and verify the fix.

Weeks to months

Customer onboarding

Coordinate setup, transform data, test integrations, wait for customer actions, and monitor adoption milestones.

Weeks to months

Renewal and retention

Build account-specific plans, schedule outreach, react to customer signals, and track renewal or expansion.

Days to months

Supply-chain exceptions

Detect shortages and delays, simulate alternatives, coordinate approvals and vendors, and track arrival.

Have a different workflow?

If it crosses systems, waits on events, needs controlled execution, and has a measurable outcome, it likely fits.

Map it with us
Across industries

Durable work exists everywhere.

The controls remain consistent while the policies, tools, systems, and outcomes change by industry.

Retail & commerce

  • Returns and warranties
  • Order exceptions
  • Replenishment
  • Marketplace maintenance

Financial services

  • KYC remediation
  • Fraud investigation
  • Document collection
  • Dispute management

Travel & hospitality

  • Disrupted-trip recovery
  • Refund coordination
  • Lost-baggage cases
  • Group bookings

Healthcare

  • Prior authorization
  • Referral completion
  • Claims follow-up
  • Care-gap outreach

Technology

  • Implementation onboarding
  • Incident response
  • Security remediation
  • Renewals

Telecommunications

  • Service activation
  • Outage resolution
  • Billing disputes
  • Churn prevention

Regulated workflows require the appropriate privacy, compliance, policy, and human-control configuration for their environment.

Qualification framework

Is this a long-running agent workflow?

The strongest candidates combine operational complexity with long periods of waiting—and cost the business every time a person has to remember what happens next.

A useful starting rule

If the process cannot be completed reliably in one browser session, it probably needs durable orchestration rather than another chatbot.

Work spans multiple systems and many steps.

The process spends meaningful time waiting for external events.

An agent needs a real environment for code, files, browsers, or tools.

Failures must survive restarts without losing progress.

Risky actions require policy checks or human approval.

Every action needs an evidence-backed audit trail.

Success can be expressed as a measurable business outcome.

Controlled autonomy

Long-running does not mean unbounded.

Every workstream runs inside explicit authority, budget, time, policy, and human-decision boundaries.

Time and turn budgets

Limit active time, turn count, failures, and duration per turn.

Scoped execution

Constrain compute, tools, credentials, network access, and workflow types.

Human control

Pause, steer, approve, reject, reprioritize, extend, or cancel.

Outcome evidence

Persist decisions, artifacts, approvals, actions, and measurable results.

Frequently asked questions

Built to keep working. Designed to stay controlled.

How long can a long-running agent work?+

A workstream can remain active for hours, days, or months. It does not keep compute running the entire time. Durable state and timers let it pause safely, release resources, and resume when an event, deadline, or human decision arrives.

What happens if an agent or worker fails?+

The workstream retains its durable state and applies configured retry, timeout, and recovery policies. Agent execution can rotate into a fresh sandbox using preserved context, checkpoints, and artifacts.

Can a human intervene while the work is running?+

Yes. Operators can approve, reject, pause, resume, steer, cancel, reprioritize, or extend an approved budget. Human judgment becomes an explicit workflow boundary instead of an ad hoc chat interruption.

Can multiple agents work on the same outcome?+

Yes. A workstream can fan work out to specialized agents running in separate sandboxes, wait for their structured results, reconcile conflicts, and apply policy before committing actions.

How do you prevent an agent from running without limits?+

Workstreams and sandboxes enforce time, turn, failure, cost, compute, network, credential, and tool boundaries. High-risk actions can require deterministic policy checks and human approval.

Design partner program

Give one important outcome a durable owner.

Bring us the process that takes hours or days, crosses systems, and still depends on someone remembering to push it forward. We will map the outcome, authority, approvals, integrations, and production measurement plan with you.