Approval gates
Refunds, merges and anything else you mark irreversible wait for a person.
Each agent writes its plan first, leaves a receipt for every step, and waits for a person before anything irreversible.
| State | Agent | Task | Receipt |
|---|---|---|---|
| Working | Refund request, order 88213 | 6 steps · 0:41 · $0.08 | |
| Waiting on you | Approve refund of $48.00 | 9 steps · 1:12 · $0.11 | |
| Handed off | Damaged item, photo review | 4 steps · 0:22 · $0.03 | |
| Done | Weekly churn summary | 14 steps · 3:05 · $0.21 | |
| Failed | Pull request 418, tests timed out | 7 steps · 2:40 · $0.09 | |
| Queued | Welcome sequence, 12 accounts | not started |
Sample data. Screens marked Example are samples while Lumenar is in pilot.
Agents mostly re-read. In our own coding-agent sessions, 99.2% of all tokens were input, and 88.5% of that input had been sent before. So we built the layer underneath: it keeps each session on its cache, picks a model for each step, and leaves you a receipt.
Measured on Lumenar’s own coding-agent logs.
What every agent does
Agents are useful when you can see what they did and stop what they shouldn’t do. Each of these is built in, not bolted on.
Refunds, merges and anything else you mark irreversible wait for a person.
Each step shows what it did, the tool it called and its time. Each session shows steps · time · cost.
No step runs before the plan is written down. You can approve it, change it or stop it.
The trail of every session is kept in an append-only log you can export.
When a case needs judgement, the agent hands it to a named person with the context attached.
A live timer runs while an agent works, next to a Stop button that ends the session at once.
How an agent works
Every session follows the same four steps, and each one leaves something you can read afterwards.
See the agentsStep 1
The agent writes its plan first. A person can approve it or change it before any work starts.
Step 2
Each step is traced: what it read, which tool it called, how long it took.
Step 3
Anything irreversible, like a refund or a merge, waits at a gate with Approve and Redirect.
Step 4
Every session ends with a receipt: steps, time and cost. The audit log is append-only.
The session layer
Requests from one agent session go to the GPU that already holds its KV cache, so old context isn’t computed again.
KV cacheSession routing
Cost of one 60-turn task
A model of how the cost falls as each mechanism is added. Not a measurement.
Inferstack prototype
Inferstack is the gateway under our agents. On a laptop with no GPU, the second turn of a session already costs a fraction of the first.
Measured on the Inferstack prototype: Qwen2.5 0.5B and 1.5B on llama.cpp, CPU only.
The agents
Answers tickets, looks up orders and drafts replies.
tickets.readorders.getpolicy.search
Waits for a person: Refunds and credits
Handles invoices, failed payments and plan changes.
invoices.getpayments.retryplans.change
Waits for a person: Any refund or plan downgrade
Reads your data and writes the weekly summary.
warehouse.querydocs.searchreport.write
Waits for a person: Sending a report outside the team
Reviews pull requests and runs the tests.
repo.readtests.runpr.comment
Waits for a person: Merging or pushing to main
Sets up new accounts and sends the welcome sequence.
accounts.createemail.draftcrm.update
Waits for a person: The first email to each account
Software
The same agents, inside calm products your team uses every day, on one account.
A support inbox where an agent drafts the reply and a person sends it.
Invoices and reconciliation, with an agent that matches payments and flags the odd ones.
Pricing
Example pricing while Lumenar is in pilot. Pilot terms are agreed with each team.
$0
for 30 days
One agent on one workflow, set up with us, to see the receipts on your own work.
12 credits
per resolved session
Pay when an agent resolves a session, from one balance of Lumenar credits.
$29
per seat / month
Lumenar products like Desk and Ledger, with agents inside.
Custom
annual
Your models, your region, your terms.
Trust
Approvals
Approvals. Anything irreversible, like a refund, a merge or an email to a customer, waits at an approval gate. A named person approves or redirects it, and their name goes in the log.
Append-only audit log
Append-only audit log. Every session leaves a trail: the plan, each step, each tool call, each approval. Entries are added, never edited, and the whole log exports.
Open models, no lock-in
Open models, no lock-in. The session layer runs open models on vLLM and SGLang and speaks the OpenAI-compatible API. The session log belongs to no single model, so you can switch models mid-session.
How session logs are used
How session logs are used. A session log is the record of one agent’s work. It is used to run that session, to hand it to another model and to show you what happened. In a pilot we agree in writing what is kept, where, and for how long.
Planned, not yet earned: SOC 2 Type II · ISO/IEC 42001
Questions
Not yet. Lumenar is in pilot, so every figure on this page is example pricing. Pilot terms are agreed with each team.
One piece of work an agent finished, such as a ticket answered or a report written, and that wasn’t handed back unresolved. Sessions an agent hands to a person are not billed as resolved.
One balance your whole organisation shares. Every agent draws from it, so you see one number, not one bill per agent.
Yes. The session layer runs open models on vLLM or SGLang and speaks the OpenAI-compatible API.
A pilot is one agent, set up with us, on your own work. You keep the receipts.