Manafold by Alomana
Your company’s AI agent work, in one record.
Manafold keeps one record of every agent's work and reads it into what people and agents do, where it fails, what it costs and what it returns. Each leader opens the view they need.
- One record of agent work
- Every change tested
- ISO 27001
Brightwater Logistics
Company view · last 90 days
AI spend tied to teams and outcomes.
AI spend
$1.24M
−12%
Agent leverage
5.1×
per person-hour
Failure rate
2.1%
−0.6 pt
Unattended work
64%
+15 pt
AI spend and return, by department
Top opportunities
- Route document work to a lighter model$210k a year
- Roll out the parallel-agent playbook to 3 teams+1,900 agent hours
- Fix 12 recurring tool failures−40% retries
Illustrative data
Enterprises running operators on Alomana
Four returns, each with its measure.
Lower AI spend
The right model for each kind of work. Waste found and stopped.
Spend per outcomeAgent leverage
People run more agents, unattended, with the playbooks of the people who already do.
Agent hours per personBetter agents
Recurring failures become tested fixes a person approves.
Same agent, then and nowBetter decisions
Vendors, seats and next steps chosen on evidence.
Licence cost and adoption
What the record answers
Read again every day, from the work itself.
- What the work isThe kinds of work people hand to agents, named from the sessions
- Where it failsRecurring failures tracked until they stay fixed
- What it costsSpend per person, team, tool and kind of work
- How far people delegateAgent hours per person, agents at once, unattended work
- What changedBefore and after a change, tested against chance
- What it returnsSpend and time against the work that shipped
Date assertions fail across a clock change
14 times by 6 people in 14 days
- FoundA recurring failure in the record
- DraftedA lesson for the agents
- TestedReplayed on past sessions where it failed
- ApprovedA person says yes
- MeasuredKept only if the failure falls
- Lesson
- Before asserting on dates in billing tests, use the account's time zone.
- Before
- 14 hits
- After
- Measured on the next weeks' sessions
- Decision
- Kept, or retired if it does not help
Writing a rule down is not enough: in our own study, rules written into an agent's instructions did not measurably cut the corrections they named.
Measured on our own work.
Token counts match each tool's own records: 97.7–100% for Claude Code, 96.9–100% for Codex.
Our team's traces, two machines, 541 Claude Code sessions.
Session time matches the tool's own turn time in 97% of turns.
One snapshot of our record.
Agents worked 757 hours and waited on a person for 1,531.
Our own team's record: about 9,800 turns, 2 people.
47 recurring issues held 72% of the failed calls that mattered.
One snapshot of our record.
On our platform, the top 1% of runs carried 67% of model spend.
30 days, 12,362 traces, list prices.
A prompt edit that looked cheaper broke 3 of 10 cases and raised cost 62%; checks on tools and side effects caught it.
Test run on a sample support agent.
How it starts
- Week 1
Connect
Plug in the agents people already use.
- Weeks 2–4
Baseline
The first map of work, cost and quality.
- Every quarter
Return
Act on the levers, measure the return.
Your people decide.
- Every change to an agent is tested on past work and approved by a person
- Secrets are redacted before anything leaves the record
- Any session can be deleted for good
Works with the agents people already use
- In the record today
- Connected as part of your rollout
- Exports to
- Trusted for
Questions
Which agents does Manafold read?
Claude Code, Codex, Gemini CLI and Hermes today, plus anything that sends OpenTelemetry, your own agents included. The rest of your stack (Cursor, Microsoft 365, ChatGPT Enterprise and others) is connected as part of the rollout.
How do you prove the return?
By counting spend and agent time against the work that shipped, before and after a change, with the assumptions on the page. The company view and the worked example here are illustrative; your baseline in weeks 2–4 replaces them with your numbers.
Do the agents change without approval?
No. A recurring failure becomes a drafted fix, the fix is tested on past sessions where it failed, and a person approves it. It is kept only if the failure falls.
Who can see what people typed?
The record keeps whole sessions, so access is by sign-in for the people you choose. Secrets are redacted before anything leaves, and any session can be deleted for good.
Where does the data live?
In a Google Cloud project and region you choose. Alomana is ISO 27001 certified, and your data is never used to train models.
How long until we see something?
Week 1 connects the agents people use. Weeks 2–4 give the first map of work, cost and quality. The return is measured every quarter after that.
See your company's AI work in one record.
Start with one team. The baseline comes in weeks.
Manafold is Alomana's record of AI agent work: what people and agents do, where it fails, what it costs and what it returns. Alomana is a secure enterprise AI platform where autonomous agents, called operators, run business processes end to end.