Feed the factory the decisions nobody wrote down
In most organizations the knowledge that matters most is decided out loud in meetings and never written. Five concrete steps route it into the context layer the factory reads.
Written by
-
Benedikt Stemmildt Founder & Co-CEO
Published: 21 August 2026 Updated: 2 min read
From our experience: in a process workshop with a client this year, a lead product owner said something that reframed the whole session. She wanted an agent that listens everywhere, summarizes features, flags inconsistencies and catches change requests. The room went quiet, because everyone knew why she wanted it.
The real knowledge in her organization was not written down. It lived on what a German colleague of mine calls the Tonspur, the audio track. Spoken in meetings, decided out loud, and never landing in a document anyone could read later.
Why a factory depends on the audio track
This matters more than it sounds, and it matters specifically for a factory.
A factory runs from ticket to pull request. The line is only as good as the context feeding it. And the context that matters most tends to live only on the verbal track: why this feature works the way it does, what got ruled out and why, which constraint is non-negotiable.
The handover meeting between a concept team and an implementation team exists for the conversation, not the slide deck. The slide deck is the part that gets saved. The conversation is the part that mattered, and it evaporates.
So the team feeds the line a clean ticket and a stale wiki, and the line produces a confident pull request built on a guess about intent. Faster, and still wrong.
Five steps to capture the verbal track
Here is the move, and it is concrete enough to start this month. Stop treating the verbal track as lost. Capture it, process it, route it into the context layer.
- Record the meetings where decisions actually get made: handovers, refinements, architecture calls. Get consent, name it openly, make it normal.
- Transcribe automatically. The transcript is raw material, not the deliverable. Nobody reads a two-hour transcript.
- Run an agent over the transcript with one job: extract the decisions. Not a summary. The specific calls: what was decided, what was ruled out, which constraint was named, who owns it.
- Have the agent flag the gaps. Where did two people assume different things? Where did a change request surface verbally and never reach a ticket? Where does what concept said differ from what implementation heard?
- Route the output into the context the line reads. The decision record goes where the agents and the humans both look, at the layer that owns that knowledge.
A line fed by what was decided
The result is a factory whose line is fed by what the organization actually decided, instead of what it remembered to type. The most reliable productivity gain I see is rarely faster code generation. It is closing the gap between what a team knows and what its line can see.
What is one decision your team made out loud this week that never reached a document?
Further reading
Written by
-
Benedikt Stemmildt · Founder & Co-CEO
More on Context engineering
- Million-token windows did not remove the dumb zone Agent quality degrades without warning above roughly 40 percent context fill, the zone Dex Horthy named. Windows grew to a million tokens and the failure mode came along, so starting bare metal is still the habit that keeps a session out of it.
- AI tool fatigue is a missing context layer Teams count their AI tools when the list that decides is the second one: skills, commands, CLAUDE.md entries and context files. Where that list is rich, the tool list shrinks on its own.
- Your coding agent is a free audit of your documentation Code generation is solved, so the constraint is now context. Every wrong answer an agent gives points at a standard the team knows and never wrote down.
Scaling needs named owners
Coding agents scale across teams when named people own the platform, the team practice, the guardrails, and the measurement. Rollout & scale is where that ownership is built and handed to internal owners, in your stack.
Training on this topic: Agentic Engineering for Developers and QAWant to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check