The AI control plane for software engineering

People lead the vision.Agents do the rest.

Your team writes down what should exist. The agents you already use plan it, price it, build it, check it, and keep watch. You say yes first and accept at the end. One set of rules for the whole company.

  • Authority

    No agent outranks the people around it

  • Money

    A ceiling before anything runs

  • Evidence

    A receipt for every delivery

WritePEOPLEPlanAGENTSBuildAGENTSCheckAGENTSWatchAGENTS
2026

The year the paradigm changed.

BeforeAn idea

NowStill yours. Write it down.

BeforeA requirements document

NowA note, in plain words.

BeforeAn estimation meeting

NowA quoted ceiling, in minutes.

BeforeSprint planning

NowThe plan comes back by itself. You say yes.

BeforeWriting the code

NowGone. The agent types.

BeforeA code review queue

NowThe agent checks its own work. You accept.

BeforeA manual test pass

NowA browser agent clicks through it.

BeforeA release checklist

NowAccepted work gathers itself. When it ships is still your call.

BeforeOn-call at 3 a.m.

NowThe agent is awake. You read it over coffee.

BeforeA status report

NowAlready written. It is the record.

BeforeTen tools, ten logins

NowOne record, one set of rules.

What is left for people

Programming is theory building.

Programming properly should be regarded as an activity by which the programmers form or achieve a certain kind of insight, a theory, of the matters at hand.

Peter Naur · Programming as Theory Building, 1985
People
1

Lead the vision

Decide what should exist, and what we will not build.

2

Set the constraints

The budget ceiling, who may touch what, and what done means.

3

Say yes

And be accountable for it.

4

Judge the evidence

Read the plan, the checks, and the receipt. Know when it is thin.

5

Hold the context

The things nobody wrote down.

Agents

Plan

Read the note and the code. Write what gets built, what done means, and a ceiling.

Build

On its own branch, across every repo the change touches, with a second agent reviewing.

Check

Click through the work in a real browser. Keep the screenshots. Say what could not be verified.

Watch

Read the logs and alerts. Catch the issue first. Work the fix.

The one rule

No agent has more authority than the people around it. Nothing costs money before a person says yes. Every decision that matters comes back to a person.

How it works

Initiatives, issues, releases. A person can steer at every step.

Above the waterline: people, togetherNoteCommentReactionReplyYesAcceptWATERLINE> session open · waiting for signalsBelow the waterline: the agent’s session

Every comment and reaction is a signal. It flows into the agent’s session, the same transcript you already know from coding with an agent.

You can step in here

What is moving?

Initiative

  1. 1You write a note. Teammates comment and react.
  2. 2A plan comes back with a ceiling, set from the agent’s own calibrated estimate.
  3. 3You say yes.
  4. 4The agent builds, with a second agent reviewing.
  5. 5It checks its own work in a real browser.
  6. 6You accept. Tested pull requests and a receipt.
What is wrong?

Issue

  1. 1The agent catches the alert first and opens an issue with a severity.
  2. 2It contains the damage.
  3. 3It proposes the fix. You say yes.
  4. 4A normal build if the fix needs code. Resolved, timeline kept.
What shipped?

Release

  1. 1Accepted work gathers into a release candidate.
  2. 2You decide when it ships.
  3. 3The release reads like a post the team can react to.
Who carried it?

Leaderboard

  1. 1People, by what they were part of.
  2. 2Models, by what they return for what they cost.
  3. 3Nothing to steer here. This one only counts.

As agents get smarter, you steer less and read fewer transcripts.

Every tool call and every transcript stays in the record, for whenever you want to look.

The whole company

Many notes. Many agents. One record.

Every team writes notes and runs agents. All of it lands in one record, under the same rules.

Four questions that matter

What is moving?

Work In Progress

Every initiative, who leads it, and what phase it is in.

What is wrong?

Issues

Incidents with a severity, an owner, and a timeline.

What shipped?

Releases

Accepted work, gathered and released by a person.

Who carried it?

Leaderboard

People by what they were part of. Models by what they return for what they cost.

Activities: One timeline of the turning points: started, approved, fixed, shipped, with people’s own words next to each.

The catalog

Each product lives in a Software System.

A software catalog of its repos and components. Every session reads it, so the next note starts knowing the software.

One set of rules

Authority, money, evidence. The same for every person and every agent.

Enforced on the server, never by a prompt.

Authority

Enterprise IAM is built in.

  • Least-privilege roles you define, and groups
  • Access scoped to exact Systems
  • Temporary access with a reason and an expiry
  • Service accounts that can never hold human authority
  • One audit trail across app, API, and MCP

Money

Every piece of work is quoted as a ceiling before it runs.

  • Your yes locks the ceiling
  • Quotes calibrated from your own delivery history
  • New automation runs in shadow on real work before it may decide anything
  • Provider bill shown separately, never marked up

Evidence

A permanent receipt for every delivery.

  • Quoted, settled, provider cost, evidence
  • The agent that builds never judges its own work
  • Budgets stop before the line, not after
  • Retries have limits; real decisions come back to a person
Four kinds of session

Each one is a session on the same Software System.

Project Management

A requirement becomes milestones, sprints, and tickets. Each ticket becomes a plan.

Engineering

A plan becomes a yes, tested pull requests, and a receipt.

QA

A browser agent clicks through what shipped and reports only what it ran.

Ops

The agent watches logs and alerts, opens the issue, and works the fix.

Keep your agent

Keep the agent you already have.

Konteks speaks ACP, the Agent Client Protocol. The agent your engineers already run connects as it is, on a laptop or in the cloud, under the same ceilings, stop points, and receipts.

ClaudeCodexPiCursorHermes⟵ ACP ⟶Konteks

Sign in with the Claude or Codex subscription you already have, or bring Pi, Cursor, or Hermes with a key.

Your provider bills you. Never marked up.

Your own machine

The agent on my laptop, or in the cloud, signed in as me, over ACP.

Connect through MCP

Same context, same rules, same stop points, whether a person or an agent is asking.

Agents
Claude, Codex, Pi, Cursor, Hermes
VCS
GitHub, Gitea
Endpoint
https://api.konteks.io/mcp
Transport
Streamable HTTP
Auth
scoped Bearer credential, rotatable and revocable

Secrets never reach a prompt, a screenshot, or a log.

SecurityDocsPricing

Learning the new discipline

If the work changed, how we learn it changed.

Nobody was trained for this job. So we learn it.

Juniors learn to judge evidence they did not produce.

Leads learn to write constraints.

Product people learn to write the note.

Refactory, the team behind Konteks, trains for exactly this.

Readiness mapping

Roles, workflows, risks, and the first work worth doing.

Role-based training

Hands-on, for engineers, leads, product people, and reviewers.

Guided onboarding

Systems, repos, roles, and the habit of saying yes on evidence.

Coaching

Review, measure, improve, after the workshop ends.

Talk to us about trainingWith Konteks, or on its own.
Last page

Unlearn the old way. Learn to work with agents.

Start building

Rolling out to a large team? Ask us about SSO/SCIM, dedicated runtimes, on-prem or VPC deployment, audit export, and onboarding.

From the journal · Day 1 · August 17, 2026

Delivery was never the hard part. The plan was.

Read