Track E / 2 volumes

Operator over OS

A later strategic track. One investment hypothesis is that workflow-specific memory, context, verification, tools and guardrails may differentiate an operator more than a shared agent platform. This is a strategic choice, not a proof or forecast of inevitable commoditisation. Counterexamples include regulated isolation, specialised devices, latency constraints and an infrastructure advantage unavailable off the shelf. Compare buy/build costs, control requirements and customer evidence before choosing. This track is coming later.

Later track All lessons are coming soon. This outline describes the planned learning path.

Volume 1

The harness and the OS

Coming soon

The durable idea: identify the requirement and compare available infrastructure with workflow-specific investment; either may be justified by evidence.

E1.01Coming soon

Models move faster than plans

Outcome: You can explain why a curriculum, a product, or a customer design should outlive any one model.

  • A six-month plan that assumes today's model is already late.
  • The four-part handoff is how you stay oriented when the model changes.
E1.02Coming soon

What an OS is for

Outcome: You can name the jobs of an agent operating system: routing, a control plane, policy, budgets, and identity.

  • These jobs are real, and someone has to do them.
  • Some platforms may bundle these functions; compare actual capabilities, costs and constraints before deciding.
E1.03Coming soon

What a harness is for

Outcome: You can name the jobs of the harness: memory, context, verification, tools, and guardrails.

  • This is where a workflow becomes trustworthy in a particular team.
  • A harness is built for a kind of work, not for a kind of model.
E1.04Coming soon

When shared infrastructure is enough

Outcome: You can argue for an OS without mistaking it for the product's lasting advantage.

  • Without routing, policy, budgets, and identity, a fleet will not stay safe.
  • With only those, you have a control plane, not an operator.
E1.05Coming soon

When workflow methods differentiate

Outcome: You can test whether workflow-specific memory, verification and guardrails create an advantage customers value, including where competitors can reproduce them.

  • A possible advantage is how reliably the work runs; test whether customers value it.
  • Customers feel the harness when a draft is right, a gate holds, and a log makes sense.
E1.06Coming soon

Operators you cannot buy off the shelf

Outcome: You can say which parts of operating a fleet have to be learned in the work, not purchased as a module.

  • Judgement about when to stop, what to hold back, and what "good" looks like in this team.
  • Those operators are people, methods, and a harness — not a licence key.
E1.07Coming soon

Memory, context, verification, tools, guardrails

Outcome: You can use those five words as a checklist for any platform, including your own.

  • If one is missing, the handoff will fail in a predictable way.
  • Track C is a memory and context story; Tracks A, B, and D are verification and guardrails.
E1.08Coming soon

Tools will change; this choice should not

Outcome: You can keep investing in the harness when the OS layer, or the model, is replaced.

  • A new OS is a change of plumbing.
  • A new harness for a kind of work is a change of product.

Volume 2

What to build, and what not to pour engineering into

Coming soon

A short investment test for a product team that wants to stay on the operator side of the line.

E2.01Coming soon

Compare infrastructure investment with buying

Outcome: You can compare a proposed build with available routing, policy and identity capabilities, including isolation, device and latency counterexamples.

  • Necessary work is not the same as differentiating work.
  • Compare buying shared infrastructure with building it; justify exceptions such as isolation or device requirements.
E2.02Coming soon

Deep workflow integration

Outcome: You can invest in steps that already live in the customer's tools, rather than in a new work surface they must adopt.

  • FusionWork's bet is that work stays in the tools the team already uses.
  • Integration quality is harness work.
E2.03Coming soon

Change management as engineering work

Outcome: You can treat briefing, adoption, and a quiet rollout as part of the design, not as a later communications task.

  • Boring use cases, a named owner, and a quiet test week are product features.
  • A brilliant harness that nobody will try has not shipped.
E2.04Coming soon

Cross-platform governance

Outcome: You can invest in policy, approval, and logs that still make sense when the customer's tools are mixed.

  • Gates and records should not be trapped in one vendor's console.
  • Cross-platform governance may justify ownership when purchased controls cannot meet the requirement; demonstrate that gap.
E2.05Coming soon

Trust, control, and operate at scale

Outcome: You can describe scale as more workflows under the same control, not as more unrestricted autonomy.

  • Least privilege, named approvers, and a readable log have to survive the second and fifth workflow.
  • Scale without those is just a wider blast radius.
E2.06Coming soon

A working system on one machine versus a slide

Outcome: You can prefer a real, running system — even a modest one — as evidence, rather than a deck about a platform.

  • A compiled second brain, or a live gated workflow, teaches a room more than an architecture diagram.
  • Demo the harness, not the logo on the OS.
E2.07Coming soon

The four-part handoff is a harness

Outcome: You can map outcome, steps in existing tools, approval gates, and an action log onto memory, context, verification, tools, and guardrails.

  • That mapping is the academy's through-line.
  • If a product idea does not strengthen one of those parts, it is probably not an initial release idea.
E2.08Coming soon

A short investment test

Outcome: You can put a proposed build through one test: does the investment meet a demonstrated workflow or infrastructure requirement at an acceptable cost?

  • If you cannot answer, do not staff it yet.
  • Revisit the test when the models move — they will.
  • Ask whether the work becomes something reusable. To count, it must work without one customer's setup, have evidence (an eval set and a baseline), and have reuse you can count. Reuse grows outward: within one customer, across a sector, then across sectors.