P/01 · The precursor

BENCH

Blueprint Engine for Normalized Context & Handoffs. The same order as SOFA, one size smaller: BENCH writes down what an agent must know before it starts, and hands it over in a form that can be checked.

Request an intro call See the mechanism

Problem

What goes wrong today.

The agent assembles its own brief

What lands in the context window is decided by the accident of the last messages. Two runs of the same task start from two different states.

The handoff is a conversation

One agent passes prose to the next. Nobody sees what actually arrived — and nobody can read it back later.

None of it leaves a trace

A context window is transient. After the run there is no longer any evidence of what the decision rested on.

Control points

What the code refuses.

Every task begins with a briefing a person has approved.: An agent that assembles its own brief has no brief. The briefing names the objective, the sources, the limits, and what is expressly not part of it.
The handoff between two agents is a file, not a chat transcript.: That makes it readable, versionable and checkable after the fact — it sits in your repository, not in the history of a session.
The briefing carries a checksum. If an input changes, the next step refuses and names the file.: A silently changed basis is the most common way a run builds something other than what was approved.

Limits

Where it does not hold.

BENCH closes no gates. It prepares.: The check is done by SOFA or by your existing review path. BENCH makes sure both are handed something checkable.
It does not replace your agent tool; it runs inside it.: That is why BENCH is the small way in: you keep your setup and gain the order that SOFA later builds on.
Without the full run there is no unbroken ledger.: BENCH records the briefing and the handoff. An unbroken record from ticket to merge is SOFA.

Delivery

What you get.

In scope

Briefing and handoff format, translated onto your storage and your terms.
Installed in a pilot repository, with the checksum lock armed.
One real task carried all the way through the precursor.
One handoff between two agents played through together.

Three phases. The first is paid and ends in a decision.

Discovery

Paid analysis phase: objective, information gaps, sources assessed technically and legally. The output is a defensible estimate — or the advice not to build it.

Build

We build the system and integrate it into your infrastructure. Fixed price based on the Discovery.

Operations

Sources and criteria change. A monthly operations agreement keeps the machine running.

An empty park bench at night after rain, close up: wet wooden slats and a cast-iron armrest in lamplight, a few autumn leaves on them, the wet pavement to the right.

Deep dive

The mechanism, in full.

Which artefacts an agentic delivery process needs, and where each of them arises, is set out in full and free of charge in our field playbook. What it does not contain is the answer for your estate: which briefing is missing at your place and which handoff breaks first. That is the paid part.

See the playbook

Six stages, the artefacts per stage, control points and measures. English and German, about 25 minutes. External page, opens in a new tab.

Contact

Bring us the question.

In discovery we establish whether it can become a system that holds. If it cannot, we say so.

Enquiry