The Bridge
Case study · CollectDev tooling
Most of my day is drafted before I see it. I click Send.
Most of my working day is answered, reviewed and drafted before I look at it. I read what the agents prepared and I approve what leaves under my name. The Bridge is the system that makes that safe. I built it to run my own practice under real client load, and it is the same system I put inside a client's company.
- What it is
- A team of AI agents that works inside a company's Slack and GitHub, drafting answers, reviews and code. A person approves anything that leaves under their name.
- Rob's role
- Designed it, built it, and runs his own practice on it every day. CollectDev's own tooling.
- Runs on
- The client's own AI subscription, cloud project, Slack workspace and GitHub. Nothing is hosted by CollectDev.
The Bridge is CollectDev's own tooling, used on client engagements. No client is named here, and the system is built so that none can be: its code names no client and its build fails if one appears. The figures below are from the live system and the window is dated so they can be refreshed. For the engineering, read how the Bridge is built.
01
AI agents that work in Slack and GitHub, with a person approving what leaves
An AI assistant for Slack that a person approves
A direct message comes in. The agents read what the team already knows, in the channels, the wiki and the repositories, and draft the reply you would have written. You read it, change it if you want to, and send it.
AI code review for pull requests
Every pull request is reviewed before anyone asks, with the verdict checked by a second agent. The review waits for you to publish it.
A ticket becomes a working branch
Hand the agents a task and they write the code on a branch of their own, answer the review comments on it, and leave the pull request for you to open.
Questions chased across three people
An errand that needs an answer from a supplier, then a colleague, then a manager moves on its own each time someone replies. Nobody has to remember to follow up.
A nightly report on what it got wrong
Every night the system grades its own day against what you actually sent, writes down what it misjudged, and turns repeat mistakes into fixes.
The agents work on their own inside that boundary. A person does one thing: approves what leaves under their name.
02
By the numbers: 28 days to 7 Nov 2026
From one client engagement, counted by the system itself.
- drafts reached a decision
- 509drafts reached a decision
- of what was sent went out exactly as drafted (166 of 168; the other 2 had edits)
- 98.8%of what was sent went out exactly as drafted (166 of 168; the other 2 had edits)
- pull request reviews delivered on 84 PRs across 4 repositories, every one sent as written
- 135pull request reviews delivered on 84 PRs across 4 repositories, every one sent as written
- nights a calibration report was written without being asked
- 27 of 28nights a calibration report was written without being asked
Code review has earned its place: every verdict in the window went out unchanged. Direct messages have not yet, so they are still drafts: Rob answered 190 of 212 himself, and every one of those was a lesson the system kept.
Zero drafts in the window were discarded as wrong.
03
Your data stays in your accounts. It runs on your own AI subscription.
Runs on your own AI subscription
The agents are Claude Code sessions, or an equivalent coding agent, on your own plan. They run in your own cloud project, read your own Slack workspace and your own GitHub. CollectDev does not host a model, does not pass your messages through anything CollectDev owns, and does not pool data across clients.
Nothing leaves without a person
Nothing the agents write reaches a colleague or a customer until a person approves it. The database enforces that rule, not a prompt: it refuses to send anything that differs from what the person approved.
No one sees more than they already could
An agent never shows anyone something they could not already see. A private conversation stays private. The system only ever brings context in, never out.
An audit trail that cannot be edited
Every action is on an audit trail that cannot be edited or deleted, including by the database owner.
One honest caveat
The AI vendor's own terms on your plan apply to what the agents send to the model, as they would for any member of staff using that plan. We do not make claims on the vendor's behalf.
The rules behind this, and how they are enforced, are on the engineering page.
04
How it gets better every day: trust earned one person at a time
Every edit a person makes to a draft is a lesson. Send it unchanged and that counts. Shorten it, warm it up, rewrite it, or answer in Slack yourself and ignore the draft: each of those is recorded against the draft and read that night.
Trust is earned, not granted. The agents start by drafting everything. A person, or a channel, is only ever promoted on a measured record of drafts that went out as written, and a single wrong draft takes that promotion back. The system never promotes itself on a feeling.
A report card is written every night without being asked. It grades the day against what the person actually did, writes what it learned into the memory the agents load next morning, and when the same mistake recurs it becomes a fix to build, not a note to remember.
05
What it costs to run
Every agent run reports what it spent, by agent and by model, and the costs page shows it by day. The expensive conversations and the expensive single runs are listed by name. One expensive night, when a report-writing agent spent millions of tokens re-reading its own inputs, changed the design: the inputs are now gathered first and the run is capped. A runaway AI bill is the first thing a buyer asks about, and here it is a number on a page, not a surprise on an invoice. How the spend is tracked.
06
What this looks like in your organization
Strip away the engineering and the Bridge is a way for people to talk to what their company already knows: the codebase, the project tracker, the ticket queue, the knowledge base, the Slack history. Someone asks in Slack. The agents read the source, answer with where the answer came from, and never show a person something they could not already see. When the answer has to go to a colleague, a person signs it. The same pattern, applied to other departments (illustrations, not deployments):
Engineering
Ask the codebase a question in Slack and get an answer with the file and line it came from. Pull requests are reviewed before anyone opens them, and tasks become branches.
The person's click: Post on a review, Open PR, Merge.
Operations
Ask the status of an order, a ticket or a project and get it from the systems, not from whoever remembers. Errands that span three people advance as each one replies.
The person's click: The commitment relayed back to the person who asked.
Finance
Ask which invoices are waiting on a PO, which contracts renew this quarter, or what was agreed with a vendor, and get the answer with its source.
The person's click: The approval, and the message that carries it.
Customer support
Each ticket triaged, the account history pulled, the knowledge base searched, the reply drafted. One customer's history never surfaces in another customer's thread.
The person's click: Send, on anything a customer reads.
Sales
Every conversation filed to the right account, and follow-ups drafted in the rep's voice with each edit scored.
The person's click: Send, on the follow-up.
People and leadership
A digest of what moved across the team's channels, grouped by theme, with personal matters left out.
The person's click: None. The digest is read only by the person it was written for.
If a department in your company spends its day answering questions that are already written down somewhere, that is the conversation to start.
07
What this shows about how we work
The Bridge is not for sale. It is the clearest evidence of what Rob Runka brings to a client's operation.
- Designing agent systems with hard boundaries: what an agent may read, what it may draft, and what only a person may release. AI and workflow automation
- Cloud infrastructure that keeps each client's data in their own project, with no shared process able to see two clients at once. How it is built
- Measurement before autonomy: trust is granted on recorded results, per person, and withdrawn the same way. The autonomy ladder
- Writing down what went wrong and fixing the cause, so the same mistake is a design change, not a reminder. Rob's experience
This is the work described under AI and workflow automation.
08
Built with
TypeScript throughout; a Fastify service and a Next.js console; Postgres on Cloud SQL with the approval rule and the audit trail enforced in the database; Cloud Run, Terraform and GitHub Actions; agents as Claude Code sessions with per-run cost metrics; Slack and GitHub or GitLab, read by the agents and written to only on a person's click. The detail.