Work arrives through email or your task system
Harness turns the incoming request into one traceable task and keeps the original message attached.
EmailTask system
How Harness works
AI does the work. Harness controls the path.
Work can arrive from a person or start automatically. Harness clarifies what needs to be done, selects the right skill and procedure, lets the AI use only the required tools, asks instead of guessing, and holds risky actions for your approval. The result returns with a complete activity record.
How it works
Two ways work enters Harness. One controlled path.
Switch between a task sent by a person and a routine started automatically. Harness controls both from start to result.
Task received01 / 06 Harness routes02 / 06 Classify the request and choose the right process
Harness identifies the expected result, permitted sources and delivery channel, then selects the matching skill and checks that the task is clear.
Controlled execution03 / 06 Run the chosen skill and process
Harness opens only the sources and actions this task needs. The AI works inside that boundary while every source and action is recorded.
Question during work04 / 06 Pause instead of guessing
If data conflicts, a source is missing or the scope would change, Harness stops at the current step.
Control point05 / 06 Validate the result and hold risky actions
Harness checks the output. Sending, changing records, deleting or signing waits for the responsible person when approval is required.
Result delivered06 / 06 Return the answer where the task came from
The report, reply, update or exception list returns to the named recipient with a readable activity record.
EmailTask system
Runs without a new message
A saved procedure runs automatically — under the same controls
The trigger, sources, skill, approval points and recipients are already defined. Harness repeats the procedure and sends exceptions to its owner.
Automatic trigger01 / 06 Start from a schedule or system event
Harness creates a traceable run at the defined time or when the connected system signals that work is ready.
Saved procedure02 / 06 Load the agreed process and its boundaries
The expected result, skill, permitted sources, stopping conditions, approval points and recipients are already set.
Permitted sources03 / 06 Collect only the data this routine may use
Harness reads the named files, mailbox, database or API and records which sources supplied the run.
Controlled execution04 / 06 Run the saved skill and handle exceptions
Normal steps continue automatically. Missing or conflicting data pauses the run and sends a precise question to the routine owner.
Control point05 / 06 Validate the result and request approval when required
Harness checks the expected output. Any sensitive action still waits for the responsible person.
Automatic delivery06 / 06 Deliver the result and keep the run record
The approved result goes to the configured recipients together with a readable activity and exception record.
EmailTask system
The AI provides reasoning. Harness owns the task, permissions, approval gates and audit trail.
The role of Harness
The AI does the reasoning. Harness keeps the work inside the rules.
An AI model can draft, analyse and decide what to try next. That does not make it safe to give the model unrestricted access to company systems. Harness gives the AI a task, exposes only the permitted capabilities, pauses defined actions for human approval, and records the run from request to result.
| The AI model | Kvantia Harness | Your team |
|---|---|---|
| Interprets the task | Restates the agreed scope | Defines the outcome |
| Plans steps | Exposes only allowed tools | Connects approved sources |
| Drafts and analyses | Enforces capability rules | Sets approval policy |
| Proposes actions | Queues sensitive actions | Approves or rejects |
| Responds to tool results | Records sources, actions and status | Reviews the result and audit |
Harness is model-independent. If you connect an external AI provider, the task content needed for reasoning is handled under that provider’s terms. Credentials, local operational state and the Harness audit trail remain outside the model context.
A concrete example
The Friday management report, without the Friday scramble
Every Friday, someone collects figures from several sources, checks them, rebuilds the same report and sends it to the same people. Here is how that routine changes with Harness.
Before
Open tools → copy figures → find a mismatch → ask a colleague → rebuild the report → draft the email → send. No single record of what happened.
With Harness
Scheduled trigger → collect permitted data → validate expected fields → flag a mismatch → build the approved template → a person approves the send → save the report with a full audit trail.
What actually changes
- Human time moves from collection and formatting to exceptions and approval.
- The routine follows the same defined sequence every week.
- A missing value becomes a visible exception, not a quiet assumption.
- The final report and the evidence behind it stay connected.
We don’t publish a numeric time-saving until it has been measured in a real pilot — a demo estimates it from your own figures.
Step by step
What happens between “do this” and “done”
- 01
Define the result
Describe what must be produced, which sources may be used, the format, the deadline and the conditions that should stop the run. A task can be one-off or saved as a recurring routine.
- 02
Confirm the interpretation
Harness restates the objective, inputs, planned output and stopping conditions. You confirm or correct the scope before the first action — an ambiguous prompt becomes an agreed work order.
- 03
Grant the minimum capabilities
Access is granted by capability. One routine may read a defined folder and create a report but never delete files or send email; another may draft an email but always require approval before sending.
- 04
Plan and perform the work
The model breaks the task into steps and uses only the capabilities Harness made available. Safe actions continue automatically; results return through Harness, so the run stays observable — Queued · Running · Waiting · Completed · Needs attention.
- 05
Stop at decisions that matter
Sending, spending, deleting, signing and any customer-defined high-impact action pause before execution. The approver sees the proposed action, its target, the reason and the evidence used.
- 06
Recover without hiding the problem
If a source is briefly unavailable, the routine retries by policy. If it cannot continue safely, it stops and says exactly where and why. Recovery never means inventing missing data.
- 07
Deliver the result with its receipt
The output arrives with a readable audit: what was requested, which sources were used, what the model proposed, which tools ran, who approved sensitive actions, what changed and when it completed.
The approval gate
Sensitive action is a queue, not a leap of faith
Sending, spending, deleting and signing pause before they run. A person sees what the AI wants to do, why, the exact target and the evidence it used.
The AI proposes an action
Harness checks the policy
Safe + allowed
Executes automatically and is recorded in the audit trail.
Approval required
A person reviews the action, target and evidence, then approves, rejects, or asks for a revision. Reject stops and records; revise returns with feedback.
The approval card shows the proposed action, its target, the reason, a data/source summary and the possible impact — with Approve, Reject and Ask for revision.
Where information goes
A clear boundary for every connection
Harness does not make every connected system “local”. It makes every connection explicit.
Stays on the Harness machine
Operational state never leaves the computer you run Harness on.
- Local policy and approvals
- Credential broker — real keys never enter the model
- Local audit trail
- Your files and operational state
Leaves only where you connect it
Relevant task content goes to what you explicitly connect — and nowhere else.
- Your chosen AI provider (e.g. Claude or GPT), for reasoning
- Business systems you connect: files, mail, databases, APIs
- Each connection is granted per capability, off by default
This website & account portal
Sees only your account, licence and device health — never your tasks, files or work.
Operational state, approval policy, credentials and the audit trail stay on the Harness machine. Relevant task content goes only to the AI providers and business systems you choose to connect.
Start small
Choose work that is easy to define and easy to verify
A good first routine is repeated weekly or monthly, follows an existing procedure, uses known sources, produces a standard output, has a clear human decision point, and costs meaningful time today.
Not a good first routine: vague strategic work; a process nobody agrees on; irreversible actions without approval; a task whose source data is unreliable; work that needs permissions the business cannot safely grant.
Before you start
A four-question readiness check
- 01
What result should exist when the routine is done?
- 02
Which exact sources are allowed?
- 03
Which actions must wait for a person?
- 04
How will we decide the result is correct enough?
Quick answers
Common questions
Does Harness replace the AI model?
No. Harness supervises the work around the model: tasks, tools, rules, approvals, recovery and evidence.
Does it need access to the whole computer?
No. Access is assigned by capability and routine. The default is no access until a capability is explicitly enabled.
Can it run while nobody is watching?
Yes, for actions the policy allows. Anything defined as sensitive waits in the approval queue.
What happens when it is unsure?
The routine can stop, request input or send an exception for review. It should not fill a missing fact with a confident guess.
Can we start with one workflow?
Yes — that is the recommended pilot: one repeated process, one owner, defined sources and a measurable before/after baseline.
See the full loop
Build the first one yourself, in an afternoon.
The setup guide takes a routine your team already repeats and walks it through scope, capability rules, the approval point, the result and the audit trail — so you can see exactly what it would do in your operation before you buy.