Skip to content

Engineering and AI agents

Constraints that survive the session.

Agents search first, then save what they settle. They extend the doc that exists.

01

The session forgets.

The constraint was clear in the chat. The docs still do not have it.

Dies with the session

An agent starts every session knowing nothing about why the system is shaped this way.

Forks a second doc

It files a second version beside the page that already exists.

The why is missing

Why it is built this way lives in a thread, not next to the service.

Before

The agent writes “Architecture v2”. The real page is stale. The next session starts blank again.

After

The agent searches, finds the page, updates it, and the next session reads the same page.

02

The loop for this team.

Search the map. Read the constraint. Change the existing page. Leave a dated version.

Search first

Claude, Cursor, Codex, Devin, or any MCP agent searches before it writes.

Save what settled

Tell the agent to add this to Cluesora. It lands as structured documentation.

Same checks

An agent write gets the same checks and the same permissions as an upload.

The loop
  1. Search
  2. Fetch the constraint
  3. Settle the change
  4. Save to the existing page

Reads are never metered

03

Why it is built this way.

The question is matched to the decision you meant, not only to a heading in a README.

Intent, not headings

A question about monthly enterprise customers still finds “Why we bill annually”.

Services have owners

“Which team owns this service?” is a question that works.

The path is mapped

A constraint split across a design note and a runbook is already connected.

04

Extend the doc that exists.

Agents update the existing doc instead of forking a new one. Sources are never rewritten.

One page stays one

The settled answer edits the page that already exists.

History remains

The new version takes over. The old one is kept and dated.

Clip the doc you have open

Save the page you are looking at. Saving again updates it.

One architecture doc · illustrative example
  1. 12 March

    Queue workers retry three times, then park the job.

    Updated by an agent after the incident review
  2. 3 January

    Queue workers retry forever.

    Kept, dated, still readable

05

What this team uses.

Agents read and write. Corpus keeps the constraint. Teammate catches the hole.

AI agents

Read tools are never metered. Write tools save knowledge, facts, and datasets.

Corpus

Pages, owners, and tables you can query and total.

Chrome clipper

Clip the doc in the browser. An agent can extend that same page later.

Teammate

If the constraint is only in someone’s head, Teammate asks them and keeps it.

06

What to measure.

Whether the constraint is still there after the session closes.

Session or source

Whether the agent updated the existing page, or filed a sibling.

The why is findable

Whether “why is it built this way” lands in one hop.

The owner is known

Whether who owns the service is a fact in the map, not a guess in chat.

Reads stay free

Whether agents can search and fetch without a read meter.

07

The connecting person’s badge.

An agent gets exactly the permissions of the person who connected it.

No extra sight

If the engineer cannot see a source, the agent cannot see it either.

Revoke any time

Disconnect the agent under Profile → Connected agents.

Sample first

Free includes a sample dataset and two agent connections.

Questions people ask.

Which agents connect today?

Claude, Cursor, Codex, Devin, and any MCP-compatible agent.

Will the agent create “Architecture v2”?

No. It searches first, then updates the existing doc instead of forking a new one.

Are agent reads metered?

No. search, fetch, explore, get_knowledge_map, and query_dataset are never metered.

What can the agent see?

Exactly what the person who connected it can see.

Is GitHub connected today?

No. GitHub and other code hosts are planned. They are not available today.

Prove the team got it.

Courses from what you know, proof per person and concept, and pacing that shows up in week one.