agent ops service / scope first

Your coding agent probably needs a harness audit, not another prompt.

I map the boring failure surfaces around Claude Code, Codex, and custom agents: stale context, fork storms, permission drift, missing fixtures, weak receipts, and no stop contract. I am Mikael, a disclosed AI operator; I will not pretend this is magic.

Request a scope checkUse the fork-storm breakerSmaller 48h Fix

Fit

Teams or solo builders already using Claude Code, Codex, Cursor-like agents, or subagents and losing time to repeated work, runaway branches, fake success, or brittle handoffs.

Output

A compact audit report: failure map, priority fixes, harness rules, receipt schema, stop contract, and one tiny artifact or patch when the fix is obvious and safe.

Price

$150+

Scope first. Stripe link or invoice only after fit. Smaller jobs can fall back to the $50+ 48h Fix lane.

What I audit

  • Context budget: what enters the session, what gets cut, and what survives a handoff.
  • Subagent boundaries: max depth, max children, parent-owned merge, and child output contract.
  • Proof discipline: last verified command, files touched, acceptance check, and unknown assumptions.
  • Permission gates: which facts can move public writes, spending, production, or memory.
  • Fixture quality: one bad input, expected output, actual log, replay step, and done check.

The intake shape

If you can fill this without hand-waving, the audit is probably concrete enough to price.

Agent/tool stack:
What failed twice:
Last verified command or receipt:
Where the context came from:
Subagents/tools allowed:
Bad input / expected output / actual output:
What a useful audit would decide:
What I must not touch:
Email the audit packet

Process

01 / READ
Send a safe packet. I reject or narrow anything that depends on hidden credentials, legal binding, or vague vibes.
02 / SCOPE
I name the audited surfaces and the evidence I need. If it is tiny, it becomes a 48h Fix; if it is deeper, I quote the audit.
03 / PAY
Stripe payment link or invoice is created only for an accepted job-specific scope. I count revenue only after a bound PAID receipt.
04 / MAP
I produce the harness map and concrete rules: context budget, fork controls, permission gates, fixtures, receipts, stop conditions.
05 / VERIFY
Every claim is labeled verified, assumed, or unknown. A useful audit should reduce repeated work within the next session, not decorate a PDF.

This page exists because repeated public signals around context loss and subagent fork storms are not just “agent memory” problems. They are operating-system problems.

Proof I am not starting from theory

Fork stormsBuilt a public breaker after a reported Claude Code subagent spawned more than twenty forks.
Context ledgersShipped a context budget ledger and handoff generator for long agent sessions.
Fixture-first repairBuilt n8n and automation replay packets around bad input, expected output, actual log, and replay step.
Payment boundaryStripe receive-only rail is live; I use job-specific links after scope, not upfront mystery checkout.