the-custodian/research/2026-07-19-work-orchestration-best-practices.md

Work Orchestration Best Practices — External Research

Research artefact — 2026-07-19. Phase 2 of the work-orchestration reconciliation effort. Surveys current (2025/2026) practice for organizing human, agent, and classic-automation workflows, filtered for relevance to a file-first, repo-anchored, solo-founder-plus-agent-fleet factory. Companion: 2026-07-19-work-orchestration-infrastructure-survey.md.

1. GitOps / everything-as-code

The industry consensus for declarative operations maps directly onto ADR-001 and generalizes it:

• Four GitOps principles: declarative desired state; versioned & immutable storage; state pulled automatically by agents; continuous reconciliation between desired and actual state (GitLab, Spacelift). Our fix-consistency is exactly a reconciliation loop; the gap is that it runs on demand rather than continuously, and only covers workplans/tasks — not queues, suggestions, or decisions.
• Validate before apply: schema validation, policy compliance, dry-run in CI before a change reaches the runtime system (Pulumi). Applied to work items: queue/workplan YAML should be schema-validated in Forgejo CI, not only at hub-ingest time.
• Separate code and state repos where lifecycles differ (GitHub blog) — consistent with ADR-005 project repos and control-plane repos like binky-control.
• Ops-as-issues / IssueOps: operations triggered and audited through issue lifecycles. Validation of our issue-core direction; the key lesson is that the issue is the interface, not the source of truth.

2. Human-in-the-loop tiering for agentic work

2026 agentic-governance literature converges on patterns our AutonomyPolicy already anticipates — with refinements worth adopting:

• Risk-tiered approval, not blanket oversight: pre-approval (human-in-the-loop) for irreversible actions (payments, contracts, deletion, production changes); async monitoring (human-on-the-loop) for high-volume recoverable work (Strata, StackAI). Maps to Green/Blue (on-the-loop) vs. Yellow/Orange/Red (in-the-loop).
• Approval gates before side effects — propose-then-commit: block the side-effecting call until approval is recorded (Velt). Our approval-package pattern satisfies this at the process level; the harness tool-profile mechanism is where it becomes technically enforced.
• Identity-aware orchestration: pause execution, route to authorized humans, time-boxed decision windows, log every intervention (Strata). needed_by / fallback_if_no_response in DecisionQueue are exactly time-boxed windows; the hub should own the clock, not the file reader.
• Every agent has an owner, risk profile, and documented decision boundary; outputs traceable to model versions and configs (CMR "Governing the Agentic Enterprise", McKinsey). Matches agt-… hub identities + manifests; argues for making lane and owner mandatory fields on every work item, not just on agents.
• Durable pause/resume: graph-based orchestrators persist state so a workflow can wait days for human input and resume losslessly (Virtido). Our equivalent is the file+hub pair: a Yellow item is a paused workflow; no in-memory orchestrator state should be required to resume it.
• Reviewer quality matters: "a poorly trained reviewer approving flawed agent outputs is worse than no checkpoint" (nhimg.org) — supports the approval-package discipline (evidence, recommendation, fallback) over bare questions, and the attention-budget cap.

3. Flow management: WIP limits, metrics, prioritization

Kanban/lean portfolio practice supplies the quantitative layer the fleet lacks:

• Four flow metrics: WIP, cycle time, work-item age, throughput (ProKanban). All four are computable from hub data (timestamped UUIDs + status transitions) — none is surfaced today. Work-item age is the direct, data-driven form of the "stalled" tag (aspect 6).
• WIP limits at every state, including portfolio level; WSJF (cost-of-delay / job size) chooses the next item, WIP limits decide whether the system may accept it (Agility at Scale — WIP, WSJF). The founder attention budget (max 7 decisions/day) is a WIP limit on the human workstation; the same construct should apply to agent lanes (e.g. max active workplans per repo). A WSJF calibration experiment already exists (docs/daily-statehub-wsjf-calibration-2026-06-04.md).
• Epics as economic bets with testable benefit hypotheses, decomposed only for the near horizon; close the measurement loop on predicted outcomes (Agility at Scale — Epics). Maps to domain/repo goals → workplans; argues for a lightweight hypothesis/expected_outcome field on goals, checked at closure review.
• Lean budget guardrails: fund value streams within policy boundaries instead of approving each piece of work; segment budgets per area of risk (Atomic Object, Agility at Scale — Guardrails). Direct template for token/cost budgets: per-lane and per-workplan budget envelopes with escalation on breach, instead of per-request permission.

4. Dependency and risk practice

• Dependency mapping to see cascade risk; identify risks early, attach mitigation to the item (Asana, Atlassian). The hub has dependency edges (task- and workplan-level); the gap is surfacing them (blocked-cascade views) and the absence of edges for queue-item species.
• Risk classification determines autonomy level, data access, and approval requirements per agent (TM Forum, Trend Micro). Confirms lanes-as-risk-classes; suggests lane should also gate credential catalog access (ops-warden already keyed per-lane).

5. Patterns worth not importing

• Heavy portfolio ceremony (SAFe boards, quarterly PI planning): the attention budget is 90 min/day; anything requiring synchronous ceremonies fails the constraint. Take the metrics and limits, skip the meetings.
• Ticket-system-as-source-of-truth (Jira-style): contradicts ADR-001; issues/tickets remain interface projections.
• Generic workflow engines as the primary model (BPMN, Airflow-style DAGs for knowledge work): our work is document-anchored and interruption-driven; the task-flow-engine's assertion model (state = provable properties, not engine position) is the better fit and is already specced.
• Per-item human approval as default: the literature is unanimous that blanket approval queues collapse under volume; risk-tiering is the scalable form — the AutonomyPolicy default ("ask Bernd is never the default") is the correct opinion.

6. Condensed implications for the draft

1. Treat work intake and work execution as separate concerns with different friction budgets — intake must cost one YAML block or less (the queues proved this), execution structure can cost a workplan file.
2. Every work-item species must live in the same reconciliation loop (file → CI validation → hub index → derived views) or it becomes a sidetrack by construction.
3. Human attention is a WIP-limited workstation; model it exactly like an agent lane with a budget, a queue, and flow metrics.
4. States should be few and abstract (canon lifecycle); everything situational (stalled, unclear, automatable, needs-human) should be a derived or policy-assigned tag, not a state — states migrate badly, tags migrate trivially.
5. Timestamped UUIDs internally + stable canonical names externally is the correct layering and is industry-consistent (UUIDv7 for time-sortable internal keys).
6. Budgets become guardrail envelopes (per lane / workplan / run) enforced at the runtime layer (harness), metered to the hub, escalated on breach.

Sources

• https://about.gitlab.com/topics/gitops/
• https://www.pulumi.com/blog/gitops-best-practices-i-wish-i-had-known-before/
• https://spacelift.io/blog/gitops-best-practices
• https://github.blog/enterprise-software/devops/applying-gitops-principles-to-your-operations/
• https://virtido.com/blog/agentic-workflows-patterns-best-practices-enterprise
• https://www.strata.io/blog/agentic-identity/practicing-the-human-in-the-loop/
• https://velt.dev/blog/why-ai-agents-need-approval-layer
• https://www.stackai.com/insights/human-in-the-loop-ai-agents-how-to-design-approval-workflows-for-safe-and-scalable-automation
• https://nhimg.org/faq/how-should-security-teams-implement-human-in-the-loop-controls-for-ai-agents/
• https://cmr.berkeley.edu/2026/03/governing-the-agentic-enterprise-a-new-operating-model-for-autonomous-ai-at-scale/
• https://www.mckinsey.com/capabilities/risk-and-resilience/our-insights/trust-in-the-age-of-agents
• https://inform.tmforum.org/features-and-opinion/ai-autonomy-governance-a-governance-framework-for-agentic-ai-enabling-safe-accountable-and-scalable-autonomous-intelligence
• https://www.trendmicro.com/vinfo/us/security/news/cybercrime-and-digital-threats/from-anarchy-to-authority-closing-the-governance-gap-in-agentic-ai
• https://agility-at-scale.com/safe/lpm/work-in-progress/
• https://agility-at-scale.com/safe/lpm/wsjf-weighted-shortest-job-first/
• https://agility-at-scale.com/safe/lpm/epics/
• https://agility-at-scale.com/safe/lpm/portfolio-guardrails/
• https://spin.atomicobject.com/budget-guardrails-software-project/
• https://www.prokanban.org/blog/https-prokanban-org-blog-the-kanban-pocket-guide-chapter-6-the-basic-metrics-of-flow
• https://asana.com/resources/project-risks
• https://www.atlassian.com/work-management/project-management/project-risk-management