A product I'm building · Multi-Agent Ops

accd.dev
run your company's AI agents

Every employee gets their own AI agent. accd.dev is the layer that manages all of them: who can do what, how much they spend, which process they follow, and who signs off before anything real happens.

accd — one run
> Send the September quote to client A
# Allowed to read Sales files ✓
# Sending email outside → needs approval
# Cost: $0.17 · Department budget 62% left
→ Draft ready with 3 sources. Waiting for the manager to approve.
4
Layers
4
Roles
10→300
Employees
100%
Runs logged
~/why-it-exists

Everyone uses AI. Few companies actually manage it

The question isn't whether AI can do the work. It's that once AI is running inside a company, nobody can answer these four questions.

01

Where did the money go?

Everyone has their own account. One big invoice arrives at month end. Nobody knows which team spent what, or which task was expensive. You can't cut what you can't see.

02

What did the agent just do?

It emailed a client, edited a file, called an API. Then the trail ends. When something breaks, all you have is a screenshot.

03

How far can it read?

A shared assistant loads every document into one place and answers everyone. Payroll, contracts, customer data. Human permissions disappear the moment the question goes through AI.

04

Is it doing it the right way?

Your company has its own processes and rules. AI doesn't know any of them. It works fast, but it works its own way, not your way.

◆
How accd.dev handles it: the four things that matter most — permissions, money, documents, process — live in our own system. Swap the model or the vendor any time. The rules don't move with them.
~/four-layers

Four layers of control

Those four questions become four layers. They share one department tree and one permission table, so they don't drift apart.

① AI Management — token spend and how people perform with AI

Every AI call goes through the platform. So every unit of spend has a name, a department and a task attached.

1.1

A ledger line per run

Each run records who ran it, which department, which task, and what it cost. View by day, week or month. View by person, team or project. Every number clicks down to the original task.

1.2

Budget caps that actually stop things

Set caps for the company, each department, each project and each person. Out of budget means the work stops and reports back. Going over requires approval.

1.3

A pre-approved list of models

Admins choose which models are available and to whom. Employees pick from that list. Changing vendors is a config change.

1.4

Who is getting value from AI

Per person and per team: how much got done, how much got sent back, cost per task. Every score clicks through to the evidence behind it.

② Project Management — the projects your people own

A project holds the conversations, files, instructions and task list together. People and agents sit in the same task board.

2.1

Private and shared projects

A private project is one person's workspace. A shared project belongs to the team, and who gets in follows the permission table.

2.2

Assign an agent like you assign a person

The assignee can be an employee or an agent. Agent tasks show a clear state: queued · running · waiting for review · failed.

2.3

Finished work still needs a human sign-off

When the agent finishes, the task isn't closed. The result goes into the owner's review queue. Click Accept, or Return with a reason.

2.4

See who is carrying what

Overdue work, silent projects, tasks nobody picked up. Visible on the board instead of asking around.

③ File Management — everyone's files, under control

This is where leaks happen. The rule: filter by permission first, then search. Never search first and redact afterwards.

3.1

Files have versions and a lifecycle

Upload → process → publish to a group → replace with a new version → archive. Old versions stay retrievable.

3.2

AI reads only what you can read

An employee's question only pulls from documents that person is allowed to see. Change teams or lose access, and the agent loses it too.

3.3

Answers must cite a source

No source, no answer. Click the citation and you get the exact passage, page and document version.

3.4

Files dropped in chat count too

Files sent into a conversation are kept, parsed and summarised. Sensitive data is masked, and every time someone unmasks it, that's logged.

④ Knowledge Base — process, working methods and compliance

This layer teaches agents how your company works. Without it, AI is fast but does things its own way. With it, an agent follows process like a trained employee.

4.1

Turn your process into an SOP

A manager answers questions, the agent writes the process out. Already have documents? Drop them in — the agent reads them and only asks about the gaps. Anything unclear gets flagged for a human to decide.

4.2

Agents follow that process

The SOP doesn't sit in a folder. It's what the agent reads before working. Approving spend, sending a quote, handling a complaint — all follow the steps the company agreed on.

4.3

Compare the process to reality

The system compares how people actually work against the written SOP. Where they differ, it says so. Two options: remind the team, or update the SOP because reality makes more sense.

4.4

Compliance with evidence

Every important action records which process it followed, which version, who approved it and when. When you need to explain yourself, you have a file, not a memory.

◆
Why it's the compliance layer: the first three answer "what is it allowed to do". This one answers "how must it be done". That's what turns AI from a personal tool into an employee that follows company rules.
~/orchestration

One big job, several agents on it

This is the biggest difference from a chatbot. accd.dev splits work across multiple agents and gives each one only the access its part needs.

Take a big job, break it into small ones

Say: "Close out payment round 3 for contractor B." The system splits it: read the contract, check the measured quantities, verify the invoices, draft the memo. Each piece becomes its own task.

Give each piece to an agent

The agents run in parallel, one part each. They report back to the main agent, which assembles the finished package.

A child agent never gets more access than its parent

Each one touches only what its part needs. There's no path for it to gain extra access. Most multi-agent setups leave this open.

◆
Agents run with the requester's permissions, not the system's. If an employee can't see payroll, neither can their agent.

Costs roll up into one number

Five agents run, but the cost lands on the one parent task. You see a single figure, not five scattered lines.

Repeating work goes on a schedule

Weekly reports, receivables checks, stalled-order sweeps. Set it once and the agent runs it. Each run is still logged and still goes through approval.

Permissions by department, not by file

The permission table is department × role. Four roles: employee, department manager, executive, platform admin.

Changing permissions shows a diff before saving. Every change is recorded.

High-risk work always involves a person

Reading data, the agent does alone. Writing into a system requires a preview. Sending anything outside the company requires an approver.

Approvers are on a clock. Miss it and it escalates.

~/under-the-hood

Four things running behind every task

Nobody sees this part. It's the part that decides whether any of it can be trusted.

Each run gets a temporary permit

Before running, the agent gets a permit stating who it's acting for, which department, which tools it may use, how much it may spend, and when it expires. No passwords or API keys inside. When it expires, it's void.

The task ledger is split in two

The summary — who ran it, what task, what it cost — is visible to management. The content — the prompt and the result — belongs to the owner. A manager can read the content, but that read is logged and the employee sees it.

Approving means seeing the real thing

The approval screen shows the exact email that will be sent, the exact row that will be written. Not a vague description. Once approved, it's locked — if the agent changes its plan, it has to ask again.

⚠
There is deliberately no "approve all" button. With that button, approval becomes theatre.

Everything lands in one log

Permission grants, budget changes, approvals, rejections, sensitive data being opened. One line each, exportable by date range.

~/examples

What those four layers look like on a normal workday

Click to expand each example.

AThe CEO asks: what did AI cost this month, and what did we get?Layer 1 · AI+
Before
Export invoices from three vendors and add them up. Still no idea what the money bought.
Now
Open the spend report. Click a department to see its tasks. Click a task to see the result and who approved it.
Result
Budget decisions based on specific work, not on a feeling.
BA manager assigns an agent, then goes to a meetingLayer 2 · Project+
Assigning
In the task board, the assignee is an agent: "Review 40 Q3 contracts, list every unusual payment term."
While it runs
The task shows running. Accumulated cost appears on the card.
Sign-off
The result goes into the review queue with files and a source per contract. Accept and it's done. Return and it goes another round.
CA new hire asks about documents they can't seeLayer 3 · File+
The question
How did this year's dealer commission policy for the north change from last year?
What happens
This person can read the general policy. They cannot read the Sales team's private annex. The answer uses only what's permitted, with sources.
The key part
The answer doesn't hint that other documents exist which this person can't see.
DHandling a customer complaint by the bookLayer 4 · Compliance+
Context
The company has a five-step complaints process written in an SOP. A new hire doesn't know it yet.
What the agent does
Reads the SOP first, then walks the employee through each step. At the compensation step it stops, because the SOP requires a manager's approval.
Later
The system compares real cases against the SOP. If 8 of 10 skip a step, the manager knows immediately: either the team is cutting corners, or that step is unnecessary.
EOne big job split across four agentsOrchestration+
The request
Prepare the payment package for contractor B, round 3.
The split
Agent 1 reads the contract. Agent 2 checks measured quantities. Agent 3 verifies invoices. Agent 4 drafts the memo. All in parallel.
Access
The invoice agent only touches invoices. It cannot see the contract unless the requester has that access.
Result
One package, one cost figure, one approval from the person in charge.
~/versus

Versus giving everyone their own AI account

The common approach today is a licence per person and let people figure it out. It works, up to a point.

Dimensionaccd.devA licence per person
CostOne ledger, by person and teamMany invoices, no link to actual work
PermissionsBy department and role, preserved through AIAnyone can paste in anything
TraceabilityFull task ledger and audit logPersonal chat history, gone when they leave
Company processAgents read the SOP firstAI does it its own way
Risky actionsPreview plus a human approverThe agent acts; you find out after
Big multi-step jobsSeveral agents, access kept separateOne chat window, you do the stitching
When someone leavesAccess revoked, data stays with the companyContext walks out with the account
⚠
Plainly: accd.dev is still being built and isn't broadly available. This page describes the architecture and the decisions already settled. What runs today versus what's still on the drawing board gets updated here.