Work that arrives with its paper trail
Scoped at kickoff against stated success criteria, signed off with evidence attached. The same rhythm for a ten-minute task and a three-week programme.
The agent platform from AIEDA Technology
Give the work to a team you can actually hold accountable.
Ollie runs standing teams of named agents under a method with gates, guards and an audit trail. Scoped before it starts, checked before it counts as done, attributable after the fact — on your infrastructure, against the model you choose.
The problem
Getting a good answer is easy.
Relying on it is the hard part.
Most tooling optimises the first mile — a clever reply in a chat box. The expensive part comes after: deciding whether the answer is right, finding out what it touched, and proving any of it to someone who was not in the room.
01
Confidence is not correctness
A plausible answer and a correct one read identically. Without an independent check, the only reviewer is the person least able to spare the time.
02
Context is rebuilt every morning
Standards and hard-won corrections live in prompts that get retyped, drift apart, and vanish the moment someone opens a new window.
03
Nothing is attributable
When an automated change causes a problem weeks later, there is rarely a record of what ran, what it was allowed to do, or who approved it.
What Ollie is
The model is one component.
Ollie is the system around it.
Swap the model, swap the terminal, move the whole thing behind your own firewall — the work still behaves the same way, because the part that determines the outcome is not the model. It is everything that happens before, around and after the model is called.
Front end
Our own terminal client, or one you already use. The interface is a choice, not a dependency — the platform underneath is the same either way.
The Ollie layer
Before the call — the directives that apply to this task, the tools that are relevant and none that are not, the grounded facts with their sources, and every correction you have already made, assembled into a prepared brief.
Around the call — tool calls intercepted and checked against the declared envelope before they run; domain work routed to real procedures: cross-checks, linting, and for silicon teams timing and area verification.
After the call — an independent pass re-verifies the claim, a second agent argues against it, and what survives is attached to the sprint as evidence. The model is never the last word on whether its own answer was right.
Model abstraction
One interface over every backend. Hosted frontier models, open models, or models running on hardware you own — selected per call, without touching the work above.
Where it runs
Your infrastructure, including fully local model hosting for teams whose material cannot reach an external API at all.
WHY IT MATTERS
Language models fail in known, repeatable ways
They drift, over-claim, forget what you told them last week, and sound equally certain whether or not they are correct. Those tendencies are predictable — which means they can be engineered against. Ollie is that engineering: the accumulated corrections that turn a capable but erratic model into something a team can depend on twice in a row.
WHAT THAT BUYS YOU
No single vendor sits on your critical path
Models change, prices change, terms change, and the best model for a task is not the same one month to the next. Because the platform does not assume any particular model underneath, none of that reaches the work your team has already built.
Platform
Four things it does that a chat box does not.
The capabilities below are what the layer above actually buys you, in the order you meet them.
01 · Named agents
Delegate to a colleague, not a blank box.
A standing roster with roles, standards and a memory of your history together. You brief them once, not every morning.
- Persistent roster — each agent keeps its role, its context and what you have already corrected it on.
- Fans out and folds back — agents spawn sub-agents for parallel work and reconcile the results themselves.
- Durable sessions — work survives restarts, branches and machines. Context travels with the task.
rhaOdile — coordinates, holds the plan active
platformMika — infra, pipelines, releases active
verifyTeodor — independent check pass queued
dataPriya — grounding and indexes idle
02 · Gates
Agents file for approval. They do not interrupt you.
Every unit of work passes two gates. Nothing executes while one is pending, and nothing pages you at midnight to ask.
- One queue — approve, amend or reject on your schedule, in one place, with the plan in front of you.
- Stated criteria — what “done” means is written down before work starts, not negotiated afterwards.
- Evidence to close — an agent returns with proof attached and waits. Done means you said so.
objectiveMove billing onto the new ledger with no gap in the audit record
successParity on 3 closed periods · zero orphaned entries · rollback rehearsed
evidenceReconciliation diff, rollback transcript
status◷ awaiting your approval
03 · Guards
The unsafe path does not exist, rather than being forbidden.
A procedure declares its envelope up front — what it may read, write and execute, and what must still be true when it finishes.
- Fails closed — reaching outside the envelope halts the run and reports it. No partial damage to unpick.
- Independently verified — a separate pass re-checks the result before it counts.
- Adversarially reviewed — a second agent argues against the finding, so plausible-but-wrong is caught before you see it.
read
ledger/**, periods/2026-*write
out/reconcile/** scopedexecutenone
must holddebits = credits on every closed period
outcome✕ halted attempted write to
ledger/2025-Q404 · Grounding
Answers that cite a source, not a recollection.
Your material is decomposed into addressable facts, each stamped with where it came from. The agent looks it up and shows you the line.
- Provenance on every fact — follow any claim back to the document, revision and line it came from.
- Indexed where it lives — wikis, drives and trackers, inside your own boundary.
- Licensed material stays put — only derived metadata moves. The source never leaves.
askedWhat notice period applies if we terminate for convenience?
answerSixty days, in writing, and any prepaid fees are refunded pro rata.
cited MSA v4 · §11.2 Order form · line 7
not foundNo carve-out for the pilot term — flagged rather than guessed.
The method
Four parts, each solving a failure the others cannot.
Not modules you buy separately — the reason the platform behaves the same way on Friday afternoon as it did on Monday morning.
Procedures with guards, not prompts with hope
A procedure declares what it may touch and what must hold true afterwards. Step outside the envelope and it stops, rather than producing a confident wrong answer.
Facts an agent can cite, not just recall
Slow-moving sources become addressable units of truth, each stamped with its origin. Agents look the fact up instead of reconstructing it.
A message plane, not a group chat
Agents address each other through real inboxes with delivery semantics. A dropped handoff is a bug you can find, not a result you never get.
Efficiency
Looking a fact up costs a fraction of reading the file it lives in.
Most agent cost is re-reading context to recover something the system already knew. Addressable facts and stored directives remove that round trip.
9×
Fewer tokens per lookup
Resolving one field from an indexed source, against reading the source file to find the same field.
10–20×
Less re-discovery
Standards resolved from stored directives instead of being restated and re-derived each session.
0
Frontier lock-in
Guarded procedures let smaller and local models do work that otherwise demands the largest model available.
Figures come from our own instrumented runs on production engineering material and are reported de-identified. Methodology and raw numbers are available under NDA — we would rather show you the measurement than ask you to take the multiple on faith.
Governance
Autonomy you can switch on because you can switch it off.
Humans gate
Approval happens in one queue on your schedule. Agents cannot page you, escalate past you, or move through a pending gate.
Fails closed
A run that reaches outside its declared envelope stops and reports, instead of doing most of the damage and apologising.
Verified twice
An independent pass re-checks the result; a second agent then argues against it.
Replayable
Every tool call, decision and approval is recorded. Months later you can answer what ran and who signed it off.
Deployment
Your infrastructure, your storage, your choice of model.
Sovereignty is not a tier we upsell. It is how the system is built, because the people we built it for cannot send their material anywhere else.
01
On your infrastructure
Inside your own network, including air-gapped, with the models themselves hosted locally when nothing may reach an external API.
02
On storage you already have
Point Ollie at the drives, wikis and trackers you run today. Artifacts stay in open formats you can read without us.
03
Hosted by us
Same boundaries, and an exit that hands back files rather than an export ticket.
No model lock-in. A single abstraction sits between the platform and whatever answers it — a hosted frontier model, an open model, or one running on your own hardware. Selection happens per call, and swapping what is underneath does not reach the work above it.
How we work
Thirty days, on your own work, with a real decision at the end.
No pilot that runs forever. We deploy against material you bring, inside your boundary, and you decide at day thirty on evidence rather than on a demo.
- 1
Days 1–5 · Stand it up
Deployed where your material already lives. Your standards become stored directives rather than instructions somebody retypes.
- 2
Days 6–25 · Do real work
Actual tasks from your backlog, run as gated sprints. You approve each one. Nothing is theatre and nothing is a sandbox.
- 3
Day 30 · Decide
You have the artifacts, the audit trail and the evidence. Continue or stop — either is a clean outcome, and you keep what was produced.
What you keep either way. Everything produced during the engagement is yours, in open formats, on your infrastructure. There is no export ticket and no hostage data.
Who it’s for
Built for work that has to be defensible.
Engineering organisations
Parallel workstreams where every change needs its review attached.
Platform and infrastructure
Teams that must prove what an automated system touched, inside their own boundary.
Regulated and IP-sensitive work
Where material cannot leave the building and “trust the model” is not an answer.
Small teams and individuals
One person running a standing team on work that would otherwise never get done.
Who builds it
Ollie comes out of production work, not a pitch deck.
AIEDA Technology builds and runs agent infrastructure for engineering organisations. Ollie is the platform that work runs on — we use it ourselves, every day, on client engagements under NDA.
Built by operators
Every mechanism on this page exists because something broke without it. The gates, the guards and the audit trail are scar tissue, not features.
Proven in regulated work
Deployed where material cannot leave the building — licensed IP, export-controlled design data, client confidential systems.
Client names on request
Our engagements are under NDA, so we do not put logos on a website. We will talk specifics under one.
Comparison
Where Ollie sits.
A fair reading of the alternatives. Each is good at something; none is trying to make the work reviewable.
| Assistant in an app | Notes & knowledge tools | General chat agents | Ollie | |
|---|---|---|---|---|
| Primary interface | A sidebar bolted onto an app | A surface you drive yourself | A chat window | A standing team you delegate to |
| Unit of work | A single prompt | A document | A conversation | A scoped, gated sprint |
| Review model | You spot-check the output | None | You spot-check the output | Kickoff and signoff gates |
| Failure mode | Confident wrong answer ships | Nothing happens | Confident wrong answer ships | Guard trips, run fails closed |
| House rules | Restated every prompt | In your head | Restated every prompt | Stored once, resolved at run time |
| Where it runs | Vendor cloud | Local, unassisted | Vendor cloud | Your infra, your storage, or hosted |
| Model choice | Fixed by the vendor | — | Fixed by the vendor | Routed per call, local included |
| Audit trail | Partial | None | Partial | Every step, replayable |
Categories describe classes of product, not any single vendor.
See it run on your own work.
A demo uses material you bring, inside a boundary you choose. You will see the gates, the guards and the audit trail doing their job — or see where they do not fit, which is also worth knowing.
We reply from a person, not a sequence.