SuperAPI AI Directory

ai-agent-orchestration

Agents & models

View repository

Updated 10-15 20:33

About this skill

AI Agent Orchestration for Construction (2026)

Why agents now

2026 construction automation is agentic: not single prompts, but specialized agents that own a domain (estimating, scheduling, documents, QA, safety), share a common data spine (the ERP + CWICR cost bases), and are coordinated by a supervisor with human checkpoints.

Agent roles

Agent Owns Tools it calls
Estimator agent BOQ + cost CWICR search, QTO, market catalogs, costs API
Scheduler agent Time (4D) task graph, dependencies, critical path, resource leveling
Document agent Specs & contracts PDF/OCR extraction, clause NER, submittal/RFI routing
QA agent Quality validation rule packs (DIN276/NRM/GAEB), reconciliation checks
Safety agent HSE checklist generation, incident classification, regulations lookup
Supervisor agent Orchestration routes tasks, resolves conflicts, escalates to humans

Coordination patterns

Supervisor ──► Estimator ──► BOQ draft ──► human approves
    │              ▲
    ├──► Document ──► scope extracted (specs) ─┘
    ├──► Scheduler ──► draft schedule from BOQ quantities
    └──► QA ──► validate BOQ + schedule, report violations
  1. Data spine first — all agents read/write the same ERP data (BOQ, tasks, cost items); no agent keeps private state.
  2. Human checkpoints — binding numbers (prices, contracts) always pass a human gate.
  3. Deterministic validation — QA uses arithmetic and rules, not LLM judgement, for reconciliation (e.g. qty × price = cost, markup conventions).
  4. Idempotent actions — every agent action is re-runnable (the ERP import is idempotent on (code, region); use it as the model).

Guardrails

  • Never let an agent invent a price: unpriced bases stay rate 0 until a market sheet exists.
  • Confidence-scored matches below threshold go to a human.
  • Log every agent decision with its inputs (the ERP's usage ledger pattern).
  • EU AI Act (2024/1689): construction estimation assistance is low/limited risk, but keep human oversight for safety-critical decisions.

Resources