← Real buildsFavorite StaffingHiring · case file

PRD_Reply Process Agent

Parses every candidate SMS reply — confirmations, available hours, start dates, corrections — writes it back to HubSpot, keeps the conversation going, and hands qualified candidates to the qualification agent. — 1,154 tasks (1,150 completed, 4 failed) in the 30 days to 2026-08-10; 675 tasks in the 14 days to 2026-07-13

In production18 nodes in the graph6 steps drawn

No separately published per-run figure; the whole measured Favorite journey is $1.16 per candidate

measured · credit-rail measured

Inbound SMS reply arrives; branch on with/without-resume path

AI parses the reply: confirmations, hours, start date, corrections

Update the HubSpot contact with what the reply said

Send the last-availability SMS (SMS 3) or an additional chase for no-resume candidates

Trigger the qualification agent when the reply completes the picture

Trigger the scoring workflow on every exit path

Where the money goes, per run

Step type	Count	Rate	Subtotal
Code / integration steps	13	$0.025	$0.325
Model calls · standard tier	2	$0.1	$0.2
Other	3	—	$0
Rate-card estimate, standard tier			$0.525

Whole graph: 13 integration nodes (5x hubspot-update-contact, 2x CustomApiTool_TriggerQualificationAgent, 3x CustomApiTool_TriggerScoringWorkflow, 2x CustomApiTool_SendSMS3, 1x CustomApiTool_AdditionalSMSforNoResume), 2 std-tier prompt nodes (Reply Handling 25,194-ch; Reply Handling with no resume 25,607-ch), 0 code, 0 parse, plus entry + 2 condition nodes. Per reply run exactly one Reply Handling fires, then the Condition routes on next_action: intake $0.025 + 1 model_std $0.10 + Update Contact $0.025 + either {Send SMS 3 + Trigger Scoring = $0.05} or {Trigger Qualification = $0.025} -> $0.175–$0.20 per run (1.75–2.0 credits). No separately measured per-run figure exists (catalog cost_measured says so); cross-check against the measured journey: $1.16 = 11.59 credits per candidate across 4 agents / 6 executions, minus screening 4.0, qualification 2.5 and $0.15 (1.5 credits) of re-entry triggers leaves ~3.6 credits for ~2 reply executions ≈ 1.8 credits each — inside the 1.75–2.0 rate-card band at std tier. No tier gap; a mid-tier assumption on Reply Handling would push a run to $0.275–$0.30 and break the journey arithmetic.

Measured: No separately published per-run figure; the whole measured Favorite journey is $1.16 per candidate (11.59 credits across 4 agents, 6 executions), of which $0.15 (13%) is re-entry triggers between the split agents (credit-rail measured)

Models seen in the graph: preferredModel GEMINI_3_FLASH on all tool nodes (fallbacks GEMINI_3_PRO/GPT4_1/CLAUDE_4_SONNET, GPT4_1_MINI on Reply Handling); resume-branch conditionNode on GEMINI_3_FLASH, but the next_action routing Condition node is pinned to GPT40 — the only non-Gemini primary in the three graphs

Graph read: ama-solution/04-agents/favorite-staffing/prd-reply-process-agent/agent.json (18 nodes, graph.agent.name 'PRD_Reply Process Agent'; flags disregarded per probe rule). Variants: 04-agents/favorite-staffing-us/reply-process-and-job-matching-favorite/ and 04-agents/favorite-staffing-eu-public/2-reply-process-and-job-matching-agent/.

What a copycat build should know

One model call per run does all the parsing: a single ~25k-char Reply Handling prompt digests confirmations, hours, start dates and corrections and emits a next_action variable; a Condition node then routes on literal matches ('next_action = send_sms_3', 'next_action = trigger_qualification_agent'). 13 of 18 nodes are pure writes/triggers.
The resume/no-resume duplication from the screening agent continues here (mirrored Reply Handling, SMS-3, Update Contact chains) — prompt changes must be applied to both 25k-char twins, which drift risk makes the priciest maintenance item.
Cross-agent choreography is webhook-based: every terminal path ends in Trigger Scoring or Trigger Qualification custom API calls — this is exactly where the journey's 13% re-entry overhead is created, and where a merged single-agent build would save $0.15/candidate.
The routing Condition still pinned to GPT40 amid an otherwise all-GEMINI_3_FLASH graph is a legacy model pin a rebuild should normalize.

Lessons from the client record

The no-resume lane is the fragile one: on the 2026-07-14 record day (93 candidates absorbed) 5 of 7 reply loops came from the no-resume path (ACTIVITY_LOG.md).
Reply-loop severity was cut 57% in one week (ceiling 35 -> 15 tasks) during operations tuning — reply loops, not throughput, were the dominant health metric (2026-07-20 weekly, ACTIVITY_LOG.md).
The structural fix — the EU reply-chat rebuild (agent 52d35526) — sat 7+ weeks without cutover and kept health At Risk; rebuilds modeled as in-flight artifacts (reply-processing/p2-build/01-eu-chat-rebuild/) stall without an owner (CLIENT.md open issue #1).
5 of the 9 migrated HubSpot nodes were in this agent (cv-screening 2, qualification 2, reply-processing 5) — the reply agent carries the most write-back surface and gained the most from the 2026-08-26 direct-API migration.

Reusable fragments cut from this agent

hrReply Handling03-projects/26-hr/02-resources/module-library/07-候选人沟通-短信.md#1
hrReply Handling with no resume03-projects/26-hr/02-resources/module-library/07-候选人沟通-短信.md#2

What it is made of

Engage a candidateHubSpotSMS via custom APIcustom qualification/scoring workflow triggers

ama-solution/04-agents/favorite-staffing/prd-reply-process-agent/agent.json · solution-intelligence/content/agent-catalog.json (id favorite-reply-process; journey $1.16 / 11.59 credits) · solution-intelligence/content/menu.json (engage-a-candidate, price_per_item 0.325) · sol-favorite-staffing/ACTIVITY_LOG.md · sol-favorite-staffing/CLIENT.md · ama-solution/03-projects/26-hr/02-resources/module-library/07-候选人沟通-短信.md