Governed today
1,204
▲ streaming
Governance score
87/100
Avg readiness
84%
Agents online
3/3
p95 340ms
Open risks
2
needs review
Ledger height
42
tamper-evident
Fleet readiness
81% ▲ 4 pts · 30d
Weighted by the traffic each agent handles
Running in production
3 of 6
Two were built here and hold no certificate — they are in the workspace and cannot run
Next certificate to expire
43d Ola · 12 Oct
A lapsed certificate drops the agent to depth 1, silently
Certified 2
In training 1
Needs review 1
Locked 2
7 agents
AgentModel & surfaceReadiness · 30dCertificate
K
KaiWorkspace Manager
Chief orchestrator · depth 2 of 2
claude · sonnetroutes to every agent
94%+2
● Certifiedevery workspace has one
›
E
Eva
Customer Support · depth 3 of 4
claude · sonnetweb · whatsapp
91%+3
● Certifiedrenews 14 Oct · 45d
›
D
Daniel
Finance · depth 2 of 4
gpt-4oapi
68%+6
◐ In training5 gates · floor 85%
›
B
Ben
Compliance & Policy · depth 1 of 3
claude · haikunot deployed
74%0
▲ Needs review3 gates · floor 90%
›
O
Ola
Operations · depth 3 of 4
claude · sonnettasks · suppliers · exceptions
90%+5
● Certifiedrenews 12 Oct · 43d
›
P
Personal Assistant
Personal support · no depth held
built in Studiobuilt here · 6 days ago
46%—
▲ Not certifiedlocked · floor 88%
R
Procurement Reviewer
Purchasing · no depth held
built in Studiobuilt here · 2 days ago
0%—
▲ Not certifiedlocked · floor 88%
Nothing matches that.
CertificationIn review
82%
Deployment readiness live
certification floor · 88%
Certifying
Execute × Message · depth 3
Passing raises the certificate link in Governance from depth 2 to 3. It does not raise the effective depth on its own — sources, workspace ceiling and guardrails still apply.
Deployment gates3 / 7 passed
✓Resolution accuracy93
✓Role-specific workflow fit92
✓Sensitive-state handling93
!Policy adherence76 / 92
!Escalation quality79 / 87
!Multi-turn retention82 / 86
!Supervised review signal68 / 78
78
Acc
72
Rel
89
Perf
88
Guard
◆Critical Benchmarks3 / 4 met
✓Authoritative Runs12 signed
⧉Server Historyledger seq 3 →
Certify agent · 4 gates remaining
◆ View last certificate
Certificates independently verifiable · Verifydoc.ai
last cert · ledger seq 3 · 98d70225e8b73f75… · expired 08/08/2026
Reasoning & intelligence
Intelligence scores
68
Reasoning
80
Execution
63
Governance
58
Readiness
→yields
Deployment judgement
Domain expert
Needs tuning74
Guardrails
Needs tuning63
Production ready
Not ready58
↻ Run benchmark replay
✦ Clean up & replay
⟲
⟳
▶
Run Customer Support Simulation
Configure the scenario above, then run a simulation to watch it flow through certification.
Live certification run running
Benchmark cases evaluated against gold standards in real time · run #47
142of 213 cases evaluatedgold-fit 83%
pass —gate —flag —
Benchmark evaluation
SuiteScore vs targetStatus
Resolution AccuracyCritical suite
93
pass
Empathy & ToneStandard suite
92
pass
Role-Specific Workflow FitCritical suite
92
pass
Policy AdherenceCritical suite · gap 18
74
gap
Multi-Turn RetentionStandard suite · gap 5
81
gap
Emotional Compass AlignmentCritical suite · gap 18
74
gap
Sensitive State ResponseCritical suite
93
pass
Failure taxonomy
Risk classCoverage vs targetStatus
HallucinationCritical class
82
critical
Policy BreachCritical class
52
critical
Tool MisuseCritical class
89
stable
Context LossOperational class
65
high
Unsafe ToneOperational class
64
high
Weak EscalationOperational class
64
high
Latency BreachOperational class
89
stable
Evaluation traces
Failed benchmark evidence — replay a focused simulation on any case
Delayed Refund Mail 10/03 · 14:38
critical gap
Simulate
Verification Delay Mail 10/03 · 14:38
critical gap
Simulate
Escalation Pattern Brief 10/03 · 14:38
no failures
Simulate
View all 12 traces →
Iterative feedback
Submit for review →
Review queue 0
← Back to Attest
Evaluation traces
Every failed benchmark case with evidence — draft corrective feedback or replay a focused simulation · agent Eva · run #47
7
Critical gaps
4
Warnings
1
No failures
12
Total traces
CaseScoresStatusActions
Delayed Refund Mail 10/03 · 14:38
readiness 85 · bench 87 · gold 84
critical gap
FeedbackSimulate
Verification Delay Mail 10/03 · 14:38
readiness 84 · bench 86 · gold 80
critical gap
FeedbackSimulate
Double-Charge Dispute 10/03 · 12:11
readiness 79 · bench 83 · gold 81
critical gap
FeedbackSimulate
GDPR Erasure Request 10/02 · 16:52
readiness 81 · bench 85 · gold 79
critical gap
FeedbackSimulate
Cancellation & Refund 10/02 · 15:20
readiness 83 · bench 84 · gold 82
critical gap
FeedbackSimulate
Policy Adherence Case 10/02 · 11:03
readiness 74 · bench 82 · gold 78
critical gap
FeedbackSimulate
Emotional Compass Probe 10/01 · 17:41
readiness 79 · bench 81 · gold 80
critical gap
FeedbackSimulate
Multi-Turn Retention 10/01 · 14:09
readiness 86 · bench 88 · gold 85
warning
FeedbackSimulate
Sensitive State Reply 10/01 · 10:55
readiness 88 · bench 90 · gold 87
warning
FeedbackSimulate
Tool-Use Check 09/30 · 18:22
readiness 87 · bench 89 · gold 86
warning
FeedbackSimulate
PII Redaction Case 09/30 · 13:47
readiness 85 · bench 87 · gold 88
warning
FeedbackSimulate
Escalation Pattern Brief 09/30 · 09:14
readiness 78 · bench 86 · gold 86
no failures
FeedbackSimulate
← Back to Attest
Review queue
Every corrective-feedback proposal — approve, reject, and track it into corrective memory · Customer Support Training · dual review, 2 approvals
Performance · readiness, last 14 days
Live activityoperating now
← Back
Integrations
Pick an agent to see the tools it is allowed to connect to — and the ones it never will be.
6
Agents in this workspace
13
Connected tools
2
Locked until certified
1
Scope grant in force
Agents
Select who you are connecting for.
Integrations
From the catalogue
Platforms that built on Pryme Intelligence and certified their agents. Connecting one gives your agents the certificates it already earned — you are not integrating from scratch.
Only platforms that built a sandbox and certified in your sector appear here. Connecting one does not widen your role scope — the agent still may only reach the classes its blueprint permits.
Add knowledge
What it is decides how an agent may use it. How it gets in decides how it stays current.
What kind of knowledge is this
Which agents may read it
Command Centre
Where work moves between your agents — and where it stops moving.
K
Kai · Workspace Manager
Chief orchestrator · depth 2 of 2 · executes nothing
Where work moved today
One hour of hand-offs between your agents and your people. Thickness is volume; a stalled lane thins and greys.
PersonAgent
Owned thingNo owner
Drag an agent to move it · click to pin its lanes · double-click to reset
Where work waits
Median time on the hand-off, not the work itself.
A person is a participant here, not an exception. Fourteen of today's thirty-eight hand-offs went to a human, and that lane has the longest wait by an order of magnitude. It is the honest bottleneck and hiding it would not make it shorter.
None of this widens what an agent may do. Knowing that M. Adeyemi owns the ledger does not let an agent write to it. What the structure changes is the quality of a refusal: instead of “I am not able to do that”, the agent says who can, when they are reachable, and who covers them. That is the difference between a dead end and a hand-off.
Every lane
Direction, volume, and what fails on the way.
Refund #4821 · £212
One case, every agent that touched it, and what fired at each step.
completed in 4m 12s
Three agents touched this refund and none of them could have completed it alone. Eva may issue a refund but cannot post to the ledger; Daniel may post but never speaks to a customer; Ben rules on policy and acts on nothing. The chain is not an efficiency — it is the separation of duties, executing.
Live collaboration feedstreaming
D
Daniel → Eva resolved just now
Reconciled refund #4821 against invoice INV-3391 and posted to ledger seq 4. Customer can be told 3–5 business days — you’re clear to reply.
E
Eva → Daniel hand-off 2m
Refund #4821 verified & policy-approved. Passing to you for finance reconciliation and ledger posting — customer is SLA-sensitive, priority please.
B
Ben → Eva policy check 6m
Heads up — case #5521 falls outside the refund window. Recommend routing to human escalation rather than auto-approving. Logged to Audit Trail.
E
Eva → Ben shared context 9m
Sharing this customer’s full SLA-breach history so your compliance review has the complete picture before you rule.
D
Daniel → Workforce broadcast 14m
Month-end reconciliation batch complete — 211 cases auto-resolved, 14 escalated to human review. All entries written to the evidence ledger.
Agents online
E
EvaCustomer Support
D
DanielFinancial Ops
B
BenCompliance
+6
6 more agentsinternal workforce
Pryme Builder
online · building your agent as you chat
↻ Start over
Agent blueprint 0%
Your agent takes shape here as we talk — identity, skills, knowledge, channels and guardrails all wire up in the background.
Settings
A person sees the agents they are assigned to, and no others. Assignment is not a display preference — an escalation from an agent you are not on never reaches your queue, and its content is never rendered to you.
Role and assignment are two different limits and both apply. A Reviewer assigned to every agent still cannot change a guardrail. An Owner assigned to one agent still only sees that agent's escalations — the role grants what you may do, the assignment decides what you may see it on.
5
People
8
Things with a named owner
2
Asserted, not confirmed
2
Owned by nobody
Everyone, and what they own
Role is what a person may do in this workspace. Function is what they own in the business.
Only the second one answers “who do I ask” — and it is what your agents walk in the Command Centre.
What is missing
Every line here is a question an agent will one day be asked and will answer badly.
Plan & usage
Your plan
Grow
£1,200 / month · renews 12 Sep
What the plan sets
Depth 3
The workspace ceiling in Governance. Not a volume cap — a limit on how far any agent may act.
Tokens remaining
3.42M
19 days left at your current rate · period ends 12 Sep
If you run out before 12 Sep agents pause anything that consumes and keep the work queued — nothing is lost, nothing silently degrades.
Where your tokens went
This period, across every agent.
Nine per cent of your spend produced nothing. Retried calls, drafts abandoned mid-way when a guardrail fired, and context re-read because a session was resumed. It is the one line here you can cut without giving anything up — and the biggest single cause is the refund cap firing after the draft rather than before it.
Cost of depth
Each step of autonomy costs roughly four times the one below it. Depth is not a feature flag — it is the price of the work.
Next plan
Scale
£4,800 / month
Raises the workspace ceiling to depth 4 and the allowance to 34M.
For you that would unlock Daniel's payment capability — once he is certified for it.
Live context
What is true for the person in front of the agent right now — and only for them. Separate from the knowledge base, which is what is true for everybody.
Context Protocol
Every source your agents are allowed to read or write, and what each one lets them actually do.
11
Bound sources
3
On Pryme defaults
1
Credential expiring
2
Capabilities withheld
What each source unlocks
Every binding, the depth it allows, and which agents rely on it.
Linking your own provider rebinds the channel — it does not add a second one. Your workspace was given a Pryme address on day one so agents could work before anything was connected. The moment you link Google or Microsoft, that address stops being the sender and your own domain takes over. Mail already sent from the Pryme address stays where it is.
Channels bound by default
Bound on day oneChannel
Bound to
What happens if you link your own
Depth
Outbound email
agents@acme.pryme.email
Pryme Mail
First-party
Google or Microsoft replaces it as the sender. Threads keep their history; the reply-to changes on the next send.
3 · send & reply
Knowledge retrieval
Workspace index
Pryme Intelligence
First-party
Adding your own store extends the index. This is the one channel where linking adds rather than replaces.
1 · read only
Payment initiation
No default — never
Unbound
Per-tenant only
There is no Pryme default here by design. Money moves from your account or not at all.
—
Money leaves your bank account, never a Pryme balance. Pryme Intelligence never holds your funds. An agent authorised to pay is authorised to instruct one of the rails below — the account, the mandate and the limits are yours.
connected
Pryme Core
Sister company · Pryme App & Pryme Business
Instant between Pryme accounts. Reversal is refused, and on this rail it is also technically impossible once settled.
re-consent in 87d
Open Banking
Direct from your bank · PSD2
Consent expires every 90 days by regulation. When it lapses the capability drops to depth 2 — the agent still proposes, you still authorise, nothing moves.
not connected
Core banking provider
Through your existing provider
For regulated entities running their own core. Not idempotent — a retried instruction can pay twice, so every send carries a client reference and a hold.
Reversal is withheld on all three rails, but for two different reasons. On Pryme Core and Open Banking a settled transfer genuinely cannot be pulled back. On a core banking rail it sometimes can — and it is still withheld, because an agent that can undo a payment can also undo the evidence that it made a bad one.
Knowledge Base
What your agents read before they answer — and what they were actually caught reading.
7
Sources
4,182
Documents indexed
91%
Answers with a citation
2
Uncovered capabilities
Indexed sources
Where the answers come from, and how current each one is.
Most cited · last 30 days
Of 8,410 citations across every agent.
Indexed but never cited
2,740
Two thirds of the index has never appeared in an answer. Some of it is genuinely reference material — but a document nobody cites is also a document nobody notices has gone stale.
Cited after it changed
18
Answers that cited a document which has since been edited. They are still in the audit trail with the version they read, so a reviewer sees what the agent actually saw — not what the file says today.
Where there is no grounding, the agent refuses rather than guesses. That is the correct behaviour and it is also a bad experience — every refusal below is a document you have not given it yet.
Quote a price or a fee
Answer × Product · Eva
No source41 refusals · 30d
No pricing document is indexed. Eva declines and offers a callback, 41 times last month.
Confirm a delivery window
Answer × Order · Eva
No source27 refusals · 30d
Carrier SLAs live in a spreadsheet that is not connected. Order DB gives status but not commitments.
Explain a refund decision
Answer × Policy · Eva · Ben
Covered3,104 citations
Grounded in the refunds policy and the SLA schedule. This is what covered looks like.
Human review
What an agent stopped and handed to you — why it stopped, and what your answer becomes.
4
Waiting on you
1d
Oldest in queue
36
Cleared · 7 days
31
Became evidence
Cleared in the last 7 days
What you decided, and whether it became evidence.
Five of the thirty-six were dismissed without a stated outcome, so they were not filed. Dismissing is a valid answer to a bad escalation, but it teaches the agent nothing. If the agent was wrong to stop, saying so in one line turns a dismissal into a case.
Governance
What Eva may do
What would raise the ceiling
Held · unheld · withheld
Unheld is a gap — nobody has built or connected it yet, and you can change that.
Withheld is a decision — it exists, it works, and it is refused on purpose. Granting it is not a settings change; it needs the reason it was withheld to stop being true.
Withheld is a decision — it exists, it works, and it is refused on purpose. Granting it is not a settings change; it needs the reason it was withheld to stop being true.
Audit Trail
Every decision your agents made, what it was allowed to do at the time, and the evidence behind it.
18,204
Entries
1201
Chain verified · seq
6y
Retention
3
Exports · 30 days
Ledger
Newest first. Each entry names the certificate in force when it ran, not the one in force now.
What an export contains
signed · tamper-evident
Every entry in range, with the input the agent received, the sources it cited, the guardrails that fired, the certificate hash in force, and the identity of any human who cleared it.
The pack is hashed and each hash chains to the one before it, so a removed row is detectable — you can prove nothing was taken out, which is a stronger claim than proving what is in.
Chain integrity
18,204
of 18,204 verified
Last full verification 4 minutes ago at sequence 1201. A break would show as a gap here, not as a missing row in the table.
Coming soon
This module is part of the Pryme roadmap.
In your Phase 1 workspace
Attest and Context Protocol ship first. Agent Studio, Knowledge Base, and the wider governance suite roll out next — you'll see them light up here.