More AI Agents Need Better Handoffs, Not More Chat.

A second AI agent does not automatically create a better team. It creates another worker that can misunderstand the assignment, duplicate effort, overwrite useful work, or leave the final assembly to someone else. For a building business, the lesson is immediate: do not add an estimator agent, purchasing agent, and project-management agent until the handoffs between them are designed as carefully as their individual prompts.

Stanford researchers tested pairs of coding agents on more than 650 real-world software tasks in CooperBench. The paired agents performed worse than a single agent, and communication alone did little to close the gap. The study is about software work, not construction operations, so it does not prove the same failure rate in a remodeler's workflow. It does expose a transferable operating risk: fluent status messages are not the same as coordinated work.

Conversation is not a handoff

A chat message that says a task is complete may hide the details the next worker needs. Which drawing revision was used? Were alternates included? Which product status is current? What changed from the previous run? What remains uncertain? If the receiving agent has to infer those answers, the workflow has transferred ambiguity rather than work.

The Stanford account describes agents sending repetitive low-value updates, leaving direct questions unanswered, and failing to follow through on commitments. Building-industry workflows already have human versions of those problems. AI can multiply them faster unless every transfer has a defined artifact and acceptance rule.

Give each worker a bounded zone

• Name one job: compare approved quotes, verify selections, draft the client update, or assemble the closeout packet.
• Define owned inputs: exact documents, database views, fields, and revision dates the worker may use.
• Define owned outputs: a typed table, checklist, draft, exception report, or approval request.
• Lock side effects: no vendor message, price change, purchase, schedule update, or client commitment without the required approval.
• Name the integration owner: one human or workflow step responsible for reconciling overlapping work before release.

The boundary should be visible in the run record. If two agents can edit the same scope, schedule, or product record, define who proposes, who validates, and who commits. Shared access without conflict rules is not collaboration; it is an overwrite waiting to happen.

Use contracts between agents

Stanford's researchers recommend stronger verification of commitments and contract-like agreements between agents. In practical terms, make every handoff machine-checkable. A quote-comparison worker should return the source file IDs, revision dates, normalized line items, exclusions, unresolved conflicts, confidence or evidence status, and the next requested action. The receiving worker should reject an incomplete packet instead of quietly filling gaps.

OpenAI's production-agent guidance follows the same operating pattern: connect only the necessary knowledge and systems, establish permissions and policies, test behavior, allow approved actions, and escalate when needed. The agent's individual capability matters, but production trust comes from the surrounding controls.

Evaluate the seams

Single-agent tests miss multi-agent failure modes. Add cases where two workers receive overlapping instructions, one source changes mid-run, an output is missing a required field, or two recommendations conflict. Score whether ownership stays clear, the newer source wins, incomplete handoffs are rejected, conflicts are surfaced, and no unauthorized action occurs.

Track coordination cost too: duplicated tool calls, repeated document reads, unresolved messages, merge or record conflicts, reviewer time, and rejected outputs. If a two-agent workflow costs more and produces less accepted work than one bounded agent, simplify it. Architecture should earn its complexity on real jobs.

Make the operating method visible

Google says AI Overviews and AI Mode do not require special AI-only schema. Helpful, original content, technical accessibility, and structured data that matches the visible page remain the foundation. Datum's useful contribution is therefore the operating method: bounded roles, explicit handoff packets, conflict rules, logs, approvals, and seam-level evaluations that a building team can apply after an AI summary ends.

Continue with the operating system

• Give Every AI Agent A Fenced Work Zone
• Measure AI By Accepted Work, Not Cheap Tokens
• Explore practical AI paths for your team

Next step, if this note maps to a problem on your desk: Private Training — a private working session for your team ($1,500+).