Skip to content

AI-Assisted Work

Workflow-orchestrated subagents, scoped tooling, and human decision boundaries.

How agent work is orchestrated, and what constrains it.

At a glance
Workflow scripts orchestrate subagents · Fan-out, pipeline, verify, judge · Allowlists, permission modes, hooks · Boundaries agents do not cross
AI-assisted work runs under the same controls as everything else: a phased plan, a verification step before commit, and atomic commits so any one step can be reviewed on its own. Agents are scoped by tool allowlists, permission modes, and hooks that block dangerous commands. They are not permitted in NDA-covered or protected code bases, and secrets never enter a prompt.
Verification tiers

How much verification an agent's output has to clear before it counts as done.

Agent leads
✓Scaffolding from an existing pattern
✓Test scaffolding and regression tests
✓Container and CI configuration
✓API exploration and documentation
✓Mechanical refactors across many files
Agent assists, output verified
~Database migration scripts
~Framework security configuration
~Architecture decision drafts
~Performance-critical code paths
~State transition logic
Decision stays human
⚠Cryptographic implementations
⚠Compliance decisions in regulated environments
⚠Legacy system institutional context
⚠Framework selection (long-term risk)
⚠Cross-system integration architecture

Decisions

OrchestrationControl surfaceWhere it applies
Pattern
Deterministic Workflow scripts drive subagents: fan-out over disjoint inputs, pipeline stages when a later stage needs an earlier stage’s output
Line of thinking
Stage order lives in a script, not in a prompt, so the same run produces the same shape twice. Fan-out when the inputs are disjoint and nothing has to be read in order. Pipeline when a stage consumes the previous stage’s result. The cost is real: the orchestrator is code I have to maintain, and a fan-out of N agents costs N times the tokens whether or not the work was independent.
FAN-OUTdisjoint inputsPIPELINEtyped handoffVERIFYSYNTHESISHUMANexit gateWorkflow script (orchestrator)stage order is code, not a promptSubagent · cluster ASubagent · cluster BSubagent · cluster Neach agent carries its own model, effort and tool allowlistStage: apply changesone cluster per agent, scoped to its filesStage: build + typecheckconsumes the previous stage's typed outputAdversarial verificationre-derives every claim from primary sourcesJudge / synthesisreconciles conflicting stage outputHuman decision gatemerge, revise, or discard

A fan-out of N agents costs N times the tokens whether or not the work was independent, and the orchestrator is code that has to be maintained like any other.

1 / 9

Where the decision stays human

Four categories an agent does not get to close.

Security & cryptography
Crypto is where a plausible implementation and a correct one look identical in review. The threat model is not in the code.
The zero-knowledge messaging design on my own product work was done by hand. Agents review it for logic errors only.
Legacy system context
Business rules accumulate for years without being written down anywhere, and an agent has no way to know what it does not know.
In NDA-covered and protected code bases, agents are not permitted at all.
Architectural trade-offs
An agent can list trade-offs. It cannot weigh them against a team, a deploy pipeline, or what you will regret in three years.
Cross-system integration
SLAs, data ownership and failure modes decide these calls, and none of them show up in a diff.