Bid review, quantity takeoff, schedule risk, RFI responses, submittal routing and weekly safety reports, run as agent teams on the Swarms API.
Precast erection sequence and crane plan
Panel LP-2 feeder size vs. one-line E-101
Rebar lap length at pile cap PC-7
Curtain wall shop drawings, sheet 3 of 9
Storm inlet rim elevation vs. paving plan
Door hardware set 04 conflicts with A-601
Open by discipline
Superintendent review
2
drafts waiting for a person
issued this shift 2
Architecture
Five reviewers read the same bid package at the same time. None of them waits on another, so the review takes as long as the slowest reader, and the response carries all five findings side by side for the preconstruction manager. Watch a run move through the desk.
Every agent gets the full task: scope letters, schedule of values, the proposed schedule and the safety plan. Each one reads for its own discipline and ignores the rest.
Long scope letters go to a fast reader such as claude-haiku-5-5. The pricing judgement, where a wrong call costs money, runs on gpt-5.4. Each agent names its own model in the request.
Each reviewer is told to cite the sheet, spec section or quote line behind every finding, so the estimator can check a claim in the documents instead of trusting the agent.
Schedule risk
A schedule risk agent reads each week's schedule export, finds the activities that slipped, follows the logic ties to see what they push, and reports the float left on everything else. The scheduler still owns the schedule; the agent writes the first draft of the risk register.
Projected finish
wk 24on plan
Float remaining
3 wk
across non-critical trades
Schedule Risk: no slips this update, finish holds at week 24.
Documents in, documents out
Agents on the Swarms API read text, so each workflow starts from an export the project team already has. The right column is what comes back, and who reads it before it leaves the trailer.
| Documents in | Where they come from | What comes back | Who reviews it |
|---|---|---|---|
| Drawings, specs and sub scope letters | Bid package export | Review packet: scope gaps, quantity checks, unit price gaps, schedule and safety findings | Preconstruction manager |
| RFIs from the field | Project management system | Drafted response with sheet and spec citations, routed to one discipline | Superintendent, then design team |
| Submittals and product data | Submittal log | Review comments against the spec section and coordination drawings | Engineer of record |
| Schedule update (P6, MS Project export) | Scheduler's weekly update | Risk register entries per trade with float remaining and drivers | Project manager |
| Daily logs, toolbox talks, incident forms | Field app export | Weekly safety report, every item linked to its log entry | Safety manager |
| Proposed change orders | Owner or sub correspondence | Scope, cost and schedule assessment with a recommendation | Project manager, owner's representative |
Code
The Swarms API examples share one request shape. The last tab runs the same agents on the open-source swarms package inside your own process. Set SWARMS_API_KEY from your Swarms Cloud account before running.
import os
from swarms_client import SwarmsClient
client = SwarmsClient(api_key=os.environ["SWARMS_API_KEY"])
# A concurrent workflow: every reviewer reads the same bid package at the
# same time and the outputs come back together as one review packet.
with open("bid_package_summary.txt", encoding="utf-8") as f:
bid_package = f.read()
response = client.swarms.run(
name="Bid Package Review",
description="Parallel scope, quantity, cost, schedule and safety review of a GC bid.",
swarm_type="ConcurrentWorkflow",
task=(
"Review the attached bid package for the Riverside Clinic, a 3-story, "
"42,000 sf outpatient building. Report findings by discipline and cite the "
"sheet number, spec section or quote line every finding comes from.\n\n"
+ bid_package
),
agents=[
{
"agent_name": "Scope Checker",
"description": "Scope gaps between the drawings, specs and sub quotes.",
"system_prompt": (
"You are a preconstruction manager. Compare the subcontractor scope "
"letters with the drawing index and spec divisions. List work that is "
"in the documents but in no one's scope, and work that two subs both claim."
),
"model_name": "claude-haiku-5-5", # fast reader for long scope letters
"max_loops": 1,
},
{
"agent_name": "Quantity Takeoff",
"description": "Sanity check of quantities against the schedule of values.",
"system_prompt": (
"You are an estimator. Check the quantities in the schedule of values "
"against the dimensions and counts on the drawings. Flag any line that "
"differs by more than 5 percent and show your arithmetic."
),
"model_name": "gpt-4.1-mini",
"max_loops": 1,
},
{
"agent_name": "Cost Estimator",
"description": "Unit prices against the firm's bid history.",
"system_prompt": (
"You are a senior estimator. Compare each major unit price with the "
"three most recent comparable bids provided and explain any gap "
"above 5 percent."
),
"model_name": "gpt-5.4", # stronger model for the pricing judgement
"max_loops": 1,
},
{
"agent_name": "Schedule Risk",
"description": "Durations and logic in the proposed schedule.",
"system_prompt": (
"You are a project scheduler. Identify activities with unrealistic "
"durations, missing predecessors and the trades on the critical path."
),
"model_name": "gemini-2.5-pro",
"max_loops": 1,
},
{
"agent_name": "Safety Reviewer",
"description": "Site-specific safety plan gaps.",
"system_prompt": (
"You are a construction safety manager. List the high-risk activities "
"in this scope (work at height, excavation, cranes, hot work) and "
"whether the submitted safety plan addresses each one."
),
"model_name": "claude-sonnet-5-5",
"max_loops": 1,
},
],
max_loops=1,
)
print(response)curl -X POST https://api.swarms.world/v1/swarm/completions \
-H "x-api-key: $SWARMS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"name": "Bid Package Review",
"description": "Parallel scope, quantity, cost, schedule and safety review of a GC bid.",
"swarm_type": "ConcurrentWorkflow",
"task": "Review the bid package for the Riverside Clinic, a 3-story 42,000 sf outpatient building. Report findings by discipline and cite the sheet, spec section or quote line each finding comes from. Scope letters, schedule of values and sub quotes follow: ...",
"agents": [
{
"agent_name": "Scope Checker",
"system_prompt": "You are a preconstruction manager. Compare sub scope letters with the drawing index and spec divisions. List work no one has claimed and work two subs both claim.",
"model_name": "claude-haiku-5-5",
"max_loops": 1
},
{
"agent_name": "Quantity Takeoff",
"system_prompt": "You are an estimator. Check schedule of values quantities against the drawings and flag any line more than 5 percent off, showing the arithmetic.",
"model_name": "gpt-4.1-mini",
"max_loops": 1
},
{
"agent_name": "Cost Estimator",
"system_prompt": "You are a senior estimator. Compare major unit prices with the three most recent comparable bids and explain any gap above 5 percent.",
"model_name": "gpt-5.4",
"max_loops": 1
},
{
"agent_name": "Schedule Risk",
"system_prompt": "You are a project scheduler. Identify unrealistic durations, missing predecessors and the trades on the critical path.",
"model_name": "gemini-2.5-pro",
"max_loops": 1
},
{
"agent_name": "Safety Reviewer",
"system_prompt": "You are a construction safety manager. List high-risk activities in this scope and whether the safety plan addresses each one.",
"model_name": "claude-sonnet-5-5",
"max_loops": 1
}
],
"max_loops": 1
}'
# The response carries "output" (the five reviews) and "usage"
# (input_tokens, output_tokens, total_tokens). The run also appears in
# Swarms Cloud under the swarm name.// RFI handling as a MultiAgentRouter: a router agent reads the question and
// hands it to one discipline specialist, which drafts the response the
// superintendent reviews before it goes to the design team.
const SWARMS_API = "https://api.swarms.world/v1/swarm/completions"
type Rfi = { number: string; subject: string; question: string; sheets: string[] }
const specialist = (name: string, model: string, scope: string) => ({
agent_name: name,
description: scope,
system_prompt:
"You are the " + name.toLowerCase() + " reviewer on a general contractor's project team. " +
"Draft a response to the RFI below. Cite every sheet, detail and spec section you rely on. " +
"Mark the draft FOR SUPERINTENDENT REVIEW at the top and list any question the design team must answer.",
model_name: model,
max_loops: 1,
})
export async function draftRfiResponse(rfi: Rfi) {
const res = await fetch(SWARMS_API, {
method: "POST",
headers: {
"x-api-key": process.env.SWARMS_API_KEY!,
"Content-Type": "application/json",
},
body: JSON.stringify({
name: "RFI Desk",
description: "Route each RFI to one discipline and draft a response for review.",
swarm_type: "MultiAgentRouter",
task:
"RFI " + rfi.number + ": " + rfi.subject + "\n" +
"Referenced sheets: " + rfi.sheets.join(", ") + "\n\n" +
rfi.question,
agents: [
specialist("Structural", "claude-sonnet-5-5", "Foundations, steel, concrete, connections and load paths."),
specialist("MEP", "gpt-5.4", "Mechanical, electrical, plumbing and fire protection coordination."),
specialist("Architectural", "claude-haiku-5-5", "Finishes, partitions, doors, hardware and envelope details."),
specialist("Civil", "gemini-2.5-flash", "Grading, utilities, paving and stormwater."),
],
max_loops: 1,
}),
})
if (!res.ok) throw new Error("Swarms API " + res.status + ": " + (await res.text()))
const data = await res.json()
return { draft: data.output as string, usage: data.usage }
}
draftRfiResponse({
number: "RFI-087",
subject: "Beam penetration for 10 inch supply duct at grid C-4",
sheets: ["S-201", "S-501", "M-301"],
question:
"M-301 routes a 10 inch supply duct through W18x35 beam B-12 at grid C-4. " +
"S-501 detail 6 limits web penetrations to 6 inches. Please confirm whether a " +
"reinforced penetration is acceptable or the duct must be rerouted below the beam.",
}).then((r) => {
console.log(r.draft)
console.log(r.usage)
})import glob
import os
import requests
BASE = "https://api.swarms.world"
HEADERS = {"x-api-key": os.environ["SWARMS_API_KEY"], "Content-Type": "application/json"}
# One request body per project. The batch endpoint runs all of them and
# returns the results in a list, so a Friday job can produce every weekly
# safety report in one call and leave one log entry per project in Swarms Cloud.
AGENTS = [
{
"agent_name": "Log Reader",
"system_prompt": (
"You read construction daily logs and toolbox talk notes. Pull out every "
"hazard observation, near miss, incident and open corrective action with "
"its date and the crew involved."
),
"model_name": "claude-haiku-5-5", # cheap reader for long logs
"max_loops": 1,
},
{
"agent_name": "Hazard Extractor",
"system_prompt": (
"You are a safety coordinator. Group the items from the log reader by "
"hazard type (falls, struck-by, caught-between, electrical, excavation) "
"and note which corrective actions are still open."
),
"model_name": "gpt-4.1-mini",
"max_loops": 1,
},
{
"agent_name": "Safety Report Writer",
"system_prompt": (
"You write the weekly site safety report for the safety manager to review "
"and sign. Every statement must cite the daily log date it comes from. "
"Do not invent counts or statistics that are not in the logs."
),
"model_name": "claude-sonnet-5-5", # stronger writer for the signed report
"max_loops": 1,
},
]
def request_for(project_dir: str) -> dict:
logs = []
for path in sorted(glob.glob(os.path.join(project_dir, "daily-log-*.txt"))):
with open(path, encoding="utf-8") as f:
logs.append(f"### {os.path.basename(path)}\n{f.read()}")
return {
"name": f"Weekly Safety Report: {os.path.basename(project_dir)}",
"description": "Daily logs to a weekly safety report for review.",
"swarm_type": "SequentialWorkflow",
"task": "Write this week's safety report from the daily logs below.\n\n" + "\n\n".join(logs),
"agents": AGENTS,
"max_loops": 1,
}
batch = [request_for(d) for d in sorted(glob.glob("projects/*"))]
resp = requests.post(f"{BASE}/v1/swarm/batch/completions", headers=HEADERS, json=batch, timeout=600)
resp.raise_for_status()
for body, result in zip(batch, resp.json()):
print("=" * 72)
print(body["name"])
print(result.get("output"))
print("usage:", result.get("usage"))from swarms import Agent, SequentialWorkflow
# The same schedule risk review on the open-source swarms package, running
# in your own process. pip install swarms, then set the provider keys for the
# models you name below.
schedule_parser = Agent(
agent_name="Schedule Parser",
system_prompt=(
"You read a construction schedule exported as text (activity id, name, "
"duration, start, finish, predecessors, total float). Normalise it into a "
"table and list the activities with total float of 5 working days or less."
),
model_name="gpt-4.1-mini",
max_loops=1,
)
float_analyst = Agent(
agent_name="Float Analyst",
system_prompt=(
"You are a project scheduler. From the parsed table, trace the critical "
"path, name the trades on it, and for each low-float activity explain which "
"predecessor logic drives it and what a one-week slip would do to the finish date."
),
model_name="gemini-2.5-pro",
max_loops=1,
)
risk_writer = Agent(
agent_name="Risk Register Writer",
system_prompt=(
"You write weekly schedule risk register entries for a general contractor. "
"One entry per trade at risk: activity, driver, float remaining, proposed "
"mitigation, owner. Plain language for the project executive."
),
model_name="claude-sonnet-5-5",
max_loops=1,
)
workflow = SequentialWorkflow(
name="Schedule Risk Review",
description="Parse the schedule, find the float, write the register.",
agents=[schedule_parser, float_analyst, risk_writer],
max_loops=1,
)
with open("riverside-clinic-week-14.txt", encoding="utf-8") as f:
schedule_export = f.read()
result = workflow.run(
"Week 14 schedule update for the Riverside Clinic. Structural steel finished "
"four working days late. Review the schedule below and write the risk register.\n\n"
+ schedule_export
)
print(result)The Python SDK (pip install swarms-client) sends one request that runs the bid review as a ConcurrentWorkflow. Five reviewers read the same package at the same time on the models that suit each job, and the response carries all five reviews with their token usage.
Use cases
Each card names the Swarms API swarm type it maps to. The agents list is the roster you would send in the request.
Five agents read the same drawings, specs and sub quotes at once: one checks scope gaps, one does a quantity sanity check, one compares unit prices with past bids, one tests the schedule, one looks for safety plan gaps. The outputs come back together as one review packet.
Each incoming RFI is read by a router that sends it to the structural, MEP, architectural or civil specialist. That agent drafts a response with the sheet and spec references it relied on, and the superintendent reviews before anything goes to the design team.
A parser reads the exported P6 or MS Project schedule, a float analyst finds the activities with the least slack and the logic ties that drive them, and a writer turns that into a weekly risk register entry per trade.
Intake reads the submittal, then a spec checker and a drawing checker work in parallel on the product data against Division 23 and the coordination drawings. A reviewer agent writes the comments in the house template for the engineer of record to sign.
Daily logs, toolbox talk notes and incident forms are read by agents that pull out hazards, near misses and open corrective actions, and a writer assembles the weekly safety report with every item linked to the log it came from.
Scope, cost and schedule agents each assess a proposed change order against the contract documents on the same run, and an aggregator writes the recommendation the project manager takes to the owner's representative.
Swarms Cloud
Runs from the API land in Swarms Cloud with their inputs, each agent's output, token usage and cost. When a finding is questioned in a meeting, the project engineer can open the run and see what the agent was given and what it wrote.
Runs / min
14
▲ last 30s
p50 latency
2.3s
▼ last 30s
Tokens / min
89.5k
▲ last 30s
Spend / hour
$3.11
▲ last 30s
Every run on the Swarms API appears in Swarms Cloud with its logs, token usage, cost and context use.
Open Swarms CloudWhat preconstruction managers, project engineers and the people who run the trailer ask before they put agents on a job.
ConcurrentWorkflow when several specialists should read the same package independently and you want all their findings together. SequentialWorkflow when one agent's output feeds the next, as in parsing a schedule before analysing it. MultiAgentRouter when each incoming item (an RFI, a submittal) should go to exactly one specialist. All of them take the same agents list, so changing the pattern is a one-field change.
Swarm types on the Swarms APIThe Swarms API takes text in the task field, so most teams export what they already have: spec sections and scope letters as text, schedules as a CSV or XER export, drawing notes and schedules as extracted text, and daily logs from the field app's export. Agents can also call tools you provide, which is how a team connects a document system or a takeoff export without copying files around.
No. Every example on this page returns a draft, a finding or a recommendation, marked for review. The superintendent, project engineer or engineer of record reads it and decides what goes out. Routing an item to a discipline is the one step the router does on its own, and the log shows which specialist it chose.
Yes. The batch endpoints accept a list of requests, one per project, and return the results together. Each request is logged on its own in Swarms Cloud with its inputs, outputs, token usage and cost, so the safety manager or project executive sees one entry per project.
Each response carries a usage block with input, output and total tokens, and the run is logged in Swarms Cloud with its cost. Mixing models keeps the price down: a fast model such as claude-haiku-5-5 or gpt-4.1-mini reads the long documents, and a stronger model writes the one report a person signs. Per-model prices are on the pricing page.
Model pricingCreate a Swarms Cloud account, copy the bid review example above, and paste in a scope letter and schedule of values. The first run shows up in your logs with its cost.
New Swarms Cloud accounts start with $5 in API credits. Enterprise teams can talk to us about volume pricing, data handling and deployment options.