A weekly seller update can require several small tasks: collect accepted enquiries, reconcile completed viewings, locate attributable feedback and prepare a draft for the agent. A long-running agent could coordinate that work, but its ability to remain active is only one part of a reliable reporting process. The report still needs a defined scope, correct source records and a clear approval and delivery boundary.
This article proposes a source-linked seller-report pilot. It does not claim that Symaxx or an agency has implemented the integration or achieved a reporting improvement. The goal is a complete draft the agent can inspect, with events and unresolved evidence accounted for. It is distinct from analysing the quality of individual buyer comments: the weekly report also covers administrative activity and reporting completeness.
Treat the beta as a development reference
Source: Introducing the Agents API is OpenAI's 10 September 2026 public-beta announcement, checked in the source snapshot on 5 October. It describes a managed agent harness and infrastructure for work involving tools, files and intermediate results. It does not document an estate-agency seller-report feature or guarantee an integration's reliability, cost or suitability.
Before selecting that development route, the technical owner must check current access, supported tools, data handling and execution behaviour for the intended deployment. A beta announcement cannot establish that a particular account or connector is available. The reporting procedure below remains useful whether the team uses an agent framework, a scheduled workflow or manual collection with a bounded drafting step.
Choose a narrow pilot, such as one agent's accepted listings and one reporting period. Give the workflow read access to only the required records. Do not grant it authority to change listing status, pricing, seller instructions or public advertisements as an incidental part of preparing a report.
Define the period and accepted event population
Specify the start and end of the reporting window, timezone and inclusion rule. Use exact boundaries agreed by the agency rather than last week in a free-text prompt. Keep enquiry receipt time separate from viewing time and report-generation time; they answer different questions.
Build an accepted event list for each listing. An enquiry is not necessarily a distinct buyer, and a viewing request is not necessarily a completed viewing. Cancelled, rescheduled and completed events need distinct states. Use stable event and contact references under the agency's privacy rules, and state the counting method in the internal evidence pack.
Collect corrections received after the period separately. The agency can decide whether to revise the earlier report or mention the correction in the next one. A model should not silently rewrite history because it found a newer note. Preserve the exact accepted snapshot used for every report version.
Reconcile facts before asking for narrative
Compute event counts deterministically and compare them with the accepted records. Check duplicate source identifiers, missing states, wrong-listing matches and events that cross the selected boundary. When completeness cannot be established, the draft should say which section is incomplete and assign an owner rather than invent a zero.
Only then ask the drafting step to explain the accepted activity in plain language. No enquiries recorded in the included sources is a narrower statement than no interest in the property. A completed viewing is an event, not evidence of an offer. A promising conversation is an agent observation, not a confirmed transaction outcome.
Source: OpenAI Structured Outputs describes schema-constrained responses. A proposed report schema can require activity counts, evidence references, feedback summaries, unresolved items and an agent recommendation section. Schema conformity does not verify counts, completeness or commercial interpretation. Validate each amount and reference against the accepted evidence pack.
Reusable weekly seller-report pilot record
Use this proposed record as the output contract. It should be complete enough for an agent to review without opening unrelated systems to discover what the draft means.
Report identity: [Report ID, listing/property IDs, agent ID, reporting window, timezone and evidence snapshot version]
Source coverage: [Included enquiry and viewing systems, last successful extraction, unavailable sources and completeness result]
| Report section | Required evidence | Wording boundary |
|---|---|---|
| Enquiries | Accepted event IDs and counting rule | Do not equate events with unique qualified buyers |
| Viewings | Completed, cancelled and rescheduled records | Do not count requests as completed visits |
| Buyer feedback | Attributable response references and missing replies | Preserve disagreement and do not infer silence |
| Current administrative actions | Accepted tasks and owner decisions | Do not invent completed follow-up |
| Unresolved items | Missing records, conflicts and responsible people | State the affected section's limit |
| Agent recommendation | Explicit agent judgement and rationale | Separate recommendation from event facts |
Draft seller update: [Concise account of accepted activity, feedback and agreed next steps; include relevant limits]
Review: [Agent accepts, revises or holds the exact draft version; amendments and reason]
Delivery: [Approved recipient, approved report version, send decision, provider/message reference and verified delivery state]
Completion check: Every count reconciles to accepted events, statements retain their evidence limits, the agent approved the exact draft and delivery is separately accounted for. A prepared draft may be complete as a draft while remaining unapproved or unsent.
Use checkpoints without duplicating report actions
A long-running process should record its progress with stable run and report identifiers. Proposed stages can include collected, reconciled, drafted, review requested, approved and delivery checked. Each stage records the evidence or decision that permits the next. A saved file is not equivalent to agent approval.
Source: OpenAI function calling describes application-side execution of model-requested functions. Retrieval tools should return identifiable records; approval and send tools should enforce their own permissions. A model saying the report was sent does not establish a provider operation or recipient receipt.
On restart, use the existing run record to identify what actually completed. If a send operation returned an uncertain result, reconcile its provider reference before retrying. Otherwise a restarted agent could deliver the same report twice. If a source extraction failed, leave that coverage gap visible and avoid advancing it to reconciled merely because the rest of the report looks plausible.
Work through a normal week and incomplete coverage
For a hypothetical listing, the accepted event list contains six enquiries, three completed viewings and two feedback responses during the agreed window. The reconciler confirms those counts and references. The draft reports the activity, preserves opposing feedback where present and marks the third viewing's reply missing. The agent supplies any recommendation separately and accepts a report version before the external delivery step.
Now suppose one enquiry arrives twice with the same stable source identifier. The accepted list retains one event and records the repeated import. If a different enquiry has the same text but a different established identity, staff inspect it rather than automatically discarding it. Similar content is not sufficient proof of duplication.
In an incomplete case, the viewing source cannot be retrieved. The process can still prepare an enquiry section from accepted records, but it marks viewing coverage unresolved and holds the overall report for the agent's decision. It does not report zero viewings. If the agent authorises a limited update, the wording makes the limit explicit and the missing-source task stays open.
Test reporting evidence and operator handover
Evaluate the pilot using fixed synthetic or appropriately permissioned records covering repeated imports, corrections, period boundaries, missing sources and interrupted delivery. Ask the agent to review factual accuracy, evidence traceability and whether the update is useful to the seller. Do not substitute a vendor benchmark for this test.
Record who owns a failed run and how to resume it. The handover should identify the last completed stage, current evidence version, unresolved records and any uncertain external action. A long-running agent is useful only if a person can understand its current state and recover without reconstructing the entire execution history.
Questions about weekly seller-report agents
Does a long-running agent need to work continuously all week?
Not necessarily. A scheduled collection and drafting process may meet the need with fewer moving parts. Choose the execution pattern after defining the report, source availability and recovery requirements. The beta development reference is one option, not a requirement to make every small reporting task autonomous.
Can the report include a recommendation to change the price?
Only as an explicitly attributable agent recommendation under the agency's professional process. Enquiry and viewing counts do not establish an appropriate price by themselves. Keep the supporting event facts, agent interpretation and accepted recommendation separate so the seller can understand what is evidence and what is judgement.
What proves the seller received the report?
A generated file or successful drafting stage does not. Retain the approved recipient, exact report version and actual delivery evidence from the chosen channel. Distinguish provider acceptance, delivered status and any stronger receipt confirmation the agency can verify. Do not invent a received outcome because the agent completed its send request.
If your business needs help defining this process, explore AI agents for estate agencies, the wider AI automation services, and our custom-agent workflow guide. The agents and automation comparison and custom AI agent glossary explain the terms. To discuss your records and approval rules, get in touch.