Talks
Vibes, meet production

Agentics Foundation Meetup · Malmö

Vibes, meet production

Simon Kihlberg Wallström · OAIZ

About me

Simon Kihlberg Wallström

Previously

Head of Architecture

Now

Co-founder, building an automation platform.

The inner loop

Where Emil left off

Emil's five-step coding-agent loop showing how the codebase supports Prompt, Orient, Retrieve, Edit, and VerifyEmil's five-step coding-agent loop showing how the codebase supports Prompt, Orient, Retrieve, Edit, and Verify
Emil’s “Write Code That AI Agents Love”
  • Navigable repos. Clear boundaries, consistent names and canonical examples.
  • Executable contracts. Types, tests and generated SDKs.
  • Linter hard rules. Encode repeated review feedback as checks.
Emil presenting at the Agentics Foundation Meetup in Malmö
Emil · June 2026

When the code met reality

Every check we had was green

Day 0The receipt-link implementation shipped.
114 daysfrom implementation commit to fix.
0 testswere old enough to notice.

The failure report came from a customer trying to close the quarter.

Zoom out

Emil’s loopCoding agentPrompt · Orient · Retrieve · Edit · Verify ↺
ShipDeploy
RealityProduction
ImpactUsers
EvidenceSignals
ContextNext change
Arm-wrestling meme joining agentic engineering and good old-fashioned DevOps to create software that keeps improving over time

How we got here

I was a good human router

  1. Two+ years agoMore nights than I can count grepping logs and correlating failures across systemsManual
  2. One year agoCopy-pasting production context into coding agentsHuman router
  3. Last fallAgents get scoped access to production systemsAgent access
  4. NowProduction signals start the investigationBuilding

The evidence stack

The data lives in five different systems

The five production evidence systems used during investigations
SystemWhat it tells us
PostHogWhat the user did in the product
GrafanaLogs, metrics and traces behind the scenes
SentryErrors and issues
Application logsWorkflow runs and agent decisions
GitHubCode, pull requests and deploy history

Every vendor in this stack also sells an AI-powered product for solving its slice of the problem.

Agent access to production · One issue

One Sentry alert became a cross-system investigation

Slack alert → Sentry + Grafana + production data → diagnosis → code → GitHub

Sentry alert arriving in Slack for a Crisp list messages API error
1. Sentry raised the signal in Slack.
Coding-agent investigation correlating Sentry issues, production executions, upstream rate limits, and the affected code path
2. One prompt connected the evidence; the same run continued into code and GitHub.

Agent access to production · One week

Sixteen issues became a nine-phase megaplan

20.2Mcumulative tokens
2h 16mwall time
102files changed
+8,713/−400lines

The inversion

The agent could read production.Could production start the agent?

Production becomes the prompt

Signals start the investigation

Production triggers and the context assembled for each one
TriggerContext assembled
User reportSession, product events, logs
Sentry issueIssue, trace, release, logs
Grafana alertMetrics, logs, traces, deploys
Post-deploy checkErrors, changed code, PR intent

The self-improving system

Different kinds of improvements

AlertThe signal was wrong. Patch the alert.
CodeThe implementation was wrong. Patch the code.
KnowledgeThe interpretation was wrong. Patch the context.
UserA person needs help. Contact them.

The outer loop we are building

The investigation is only the first half

What we do today

Triage

Signal → production context → diagnosis

End-to-end automation

Pull request delivered

Signal → context → diagnosis → code → pull request

What triage looks like today

The deployment starts the investigation

Automation workflow triggered by a GitHub deployment, waiting thirty minutes, analyzing the deployed revision and Grafana logs, then posting the result to Slack
Deploy → wait 30 minutes → analyze the revision and Grafana logs → post to Slack.

What triage looks like today

The report is the current handoff

Slack post-deployment observation report correlating the workflow run, deployed commit, pull request, observation window, production logs, synthetic probes, assessment, and next action
The system verified this deployment as healthy and recommended no remediation.

Last 90 days

Agents opened almost as many PRs as humans

466autonomous
559human
PRs opened
26%of autonomous merges had no human touch

Autonomous triage → manual routing → autonomous implementation

The point

The hard part is deciding what work should start

Your turn

What stage are you at?

QR code linking to Simon Kihlberg Wallström on LinkedInLinkedInlinkedin.com/in/simon-kihlberg-wallström
Slidessimon.kihlbergwallstrom.com/talks/vibes-meet-production