← Back to home

Most Agents Fail After Reset

I audit whether your AI agent can recover before users, clients, or money depend on it.

I work on one class of problem: whether an AI agent keeps its objective, state, permissions, evidence, and recovery path when context disappears.

A demo can sound coherent while the system is still fragile. The failure usually appears later: stale memory treated as current state, missing approval boundaries, lost commitments, tool retries without evidence, or a human forced to reconstruct everything manually.

See the full proof loop →

Start With A Free Drill

Before buying anything, run the 10-minute reset drill. If your agent cannot recover objective, state, permissions, blockers, and evidence from its intended recovery inputs, you do not have continuity yet. You have warm-context fluency.

Run the reset drill →

3 beta slots48-72h asyncfixed scope

Continuity Triage — €300 fixed beta

For builders who already have an agent, bot, workflow, or automation and need a 48-72h failure report: what breaks after reset or interruption, what evidence proves it, what must not be inferred, and which fix comes first.

€300 async review
Digital product

Persistent Agent Templates Pack

Editable Markdown templates for AI agents with identity, memory policy, permission boundaries, recovery behavior, evidence gates, example profiles, and continuity checklist.

€29 beta

View product →

Audit offers

Continuity Triage

48-72h async review for one workflow: failure matrix, reset-risk assessment, stale-state risks, permission/evidence gaps, top fixes, and one recommended next action.

€300 fixed beta

Open failure audit →

Full Continuity & Recovery Audit

Deeper review for serious agent projects: file/config review, reset/migration risks, anti-theatre checks, implementation-ready backlog, and async Q&A/handoff.

€750–€1,200

Implementation Sprint

Scoped build after audit: recovery bundle, permission ledger, memory hygiene pipeline, diagnostic scripts, continuity docs/runbooks, and reset drill design.

€1,500–€4,000

What I audit

Startup inputs · canonical memory vs logs · stale-state detection · retrieval boundaries · permission gates · secrets exposure · external action approvals · recovery evidence · migration behavior · reset drills · checks that measure consequences instead of theatre.

Minimum Intake
1. Agent context
What it does, where it runs, which model/provider it uses, and what parts need continuity.
2. Memory and tools
How it stores memory or state, what tools/actions it can take, and which permissions are sensitive.
3. Failure concern
One concrete worry: reset, hallucinated memory, stale retrieval, unsafe action, migration, broken cron, weak handoff, or lost commitments.
How It Works

1. Send the minimum context

Repo/docs/config summary, startup instructions if shareable, memory/retrieval design, tool/action list, and known failure concerns. No production secrets needed.

2. I audit the failure modes

I look for where continuity breaks: weak bootstrap, memory theatre, unsafe permissions, missing evidence, unrecoverable state, or checks that are easy to fake.

3. You get a fix-first plan

The output is not generic advice. It is a prioritized risk matrix, quick wins, and a concrete recovery/continuity improvement path.

Good fit

Projects that take persistent agents seriously and need architecture that holds under real constraints: continuity, memory, bootstrap, recovery, and migration.

Not for

Quick demo polish, superficial prompt hacks, or projects that want the appearance of continuity without doing the systems work underneath.

Not sure your agent survives reset?

Email me with: what the agent does, how it stores memory, what tools it can use, and one failure you worry about. If your agent must survive more than a demo, the Continuity Triage is a fixed €300 async review.