Human Agent Delivery System

Human-Agent Delivery System

Owner: Bill Eisenhauer Last Updated: August 6, 2026 Status: 🚧 Working model Contributors: Bill Eisenhauer

---

A hybrid organization is not simply a human team with faster coding. It is a socio-technical system in which different kinds of judgment, execution, evidence, and authority can be allocated deliberately.

A reference pipeline

flowchart LR
    S["Shape\nHuman-led intent"] --> K["Full Kit\nShared context"]
    K --> B["Build\nHuman + agent execution"]
    B --> V["Verify\nAutomated evidence + judgment"]
    V --> R["Integrate / Release\nSystem ownership"]
    R --> A["Adopt / Own\nConsumer behavior"]
    A -. telemetry .-> S

Loading

In prose: people frame intent; a full kit makes context authoritative; humans and agents execute; automation and human judgment verify; system owners integrate and release; consumers adopt and own. Telemetry from the whole path informs the next shaping decision.

Allocate work by evaluability and consequence

The central design question is not “human or agent?” It is “what level of autonomy can this task safely support given how clearly success can be evaluated and how costly failure would be?”

Task shape	Suggested pattern	Human responsibility
Clear rules, strong tests, reversible effects	Agent executes inside a bounded workflow	Set constraints and maintain the evaluator
Several valid paths, observable outcome	Agent plans and acts with checkpoints	Approve intent, exceptions, and consequential actions
Ambiguous goal or weak acceptance evidence	Human shapes; agent researches or drafts	Resolve ambiguity and define success
High-consequence or hard-to-reverse change	Human-led execution with agent assistance	Retain decision authority and final acceptance

Anthropic distinguishes predefined workflows from agents that dynamically direct their own process and tool use. That distinction is operationally useful: choose the simplest architecture that fits the work, then add autonomy only when the evaluation and control system can support it.

Scarce human attention is capacity

Agent output can make several forms of human work more—not less—important:

• deciding what outcome matters;
• supplying authoritative context;
• designing tests and evaluation rubrics;
• reviewing high-risk or ambiguous changes;
• resolving cross-system integration;
• earning adoption; and
• accepting long-term ownership.

Treat these as explicit capacity pools. If they are invisible, the system will discover them through queues, rework, interruptions, and aging work.

flowchart TD
    O["More agent output"] --> Q1["Verification queue"]
    O --> Q2["Integration queue"]
    O --> Q3["Adoption queue"]
    Q1 --> H["Scarce human attention"]
    Q2 --> H
    Q3 --> H
    H --> W["WIP and delay when overloaded"]

Loading

In prose: agent output can create three downstream queues that compete for the same human attention. When that shared resource is overloaded, WIP and delay rise even if implementation is fast.

The minimum operating rig

A team does not need a large transformation program to learn. Start with:

1. an explicit start and finish boundary for value;
2. event timestamps at each important handoff;
3. WIP, throughput, work-item age, and cycle time;
4. first-pass acceptance or rework evidence;
5. human-attention estimates at scarce decisions;
6. an adoption and ownership signal; and
7. a six-week intervention log with a named hypothesis.

DORA's 2025 research describes AI as an amplifier of an organization's existing strengths and weaknesses. This is compatible with the model here: the underlying delivery system determines whether additional AI capacity becomes leverage or additional pressure.

---

Related: Kitting and Wait States · Flow Discovery and Intervention Rig · Constraint Learning Loop · Sources and Further Reading