Orzed Intake Agent, brief comprehension at the front door
- Pulse based; reads briefs in seconds and emits a structured Intake Report.
- Trained on roughly 8,000 historical client briefs paired with senior engagement lead annotations.
- Output is labelled exploratory in the Console; binding decisions belong to the Technical Review Team.
- Catches the major brief level gap or risk in roughly 80 to 88 percent of cases on the held out evaluation set.
- Does not speak to the customer directly. The Intake Report goes to the human reviewer who decides what to surface.
The first agent every Orzed engagement meets is the Intake Agent. The customer submits a brief through the Console; within seconds the Intake Agent has read it and produced a structured Intake Report. That report goes to the Technical Review Team alongside the original brief. The team reads both, decides what to clarify with the customer, and decides whether the engagement is ready to move into planning.
This is the agent card. It explains what the Intake Agent is, what it is trained on, what it produces, where its limits are, and why its output is structurally exploratory rather than binding.
What the agent does
The Intake Agent reads a brief and produces an Intake Report. The Report is a structured document with named sections; it is not free form prose.
The sections are stable across engagements. Summary restates the brief in the agent’s own words, which is the simplest and most reliable check that the agent understood what it read. Gaps lists missing information the engagement will need before planning can begin (acceptance criteria, success metrics, compliance frame, integration contracts, throughput expectations). Ambiguities lists places where the brief could be read two ways and the disambiguation will materially change scope. Risks lists technical, scope, timeline and compliance risks the agent surfaces from the brief. Domain notes flags any specialised vocabulary the agent recognises and is not confident about, prompting the human reviewer to verify domain specific assumptions.
That structure is the contract. The Technical Review Team relies on knowing where to look for what; a free form report would be slower to read and easier to skim past.
Architecture
The Intake Agent is a thin orchestration over Orzed Pulse. The Pulse model is invoked once per brief (or chained twice for very long briefs) with a prompt that defines the structured output format and a small library of in context examples drawn from past engagements. There is no tool use loop here; the agent is a single shot reader and writer, not a researcher.
This shape is deliberate. The Intake Agent’s job is to read the brief and produce a report on it, not to go fetch information from external sources. Augmenting the brief with external research is a Planning Agent task and lives at a different layer. Keeping the Intake Agent narrow keeps it fast (the report comes back in two to four seconds on a normal brief) and keeps its failure modes well understood.
Training
The fine tune draws from three sources.
Historical client briefs. Roughly 8,000 anonymised briefs from past engagements, paired with the Intake Reports a senior engagement lead would have written for them. The pairing was created retrospectively for the older briefs (a senior lead reviewed each and produced a Report) and incrementally for newer ones (the Intake Reports were captured as engagement leads reviewed live briefs over an 11 month period).
Engagement retrospectives. When an engagement closes, the team writes a retrospective covering what was missed at intake, what risks materialised, what scope shifted. These retrospectives are linked back to the original brief and provide a labelled “what should have been caught” signal that reinforces the right reading habits in the agent.
Negative examples. Briefs that were misread or misclassified in production, paired with the corrected Report. These are weighted more heavily in the loss; the agent learns most efficiently from the cases where its earlier versions failed.
The fine tune is supervised. There is no reinforcement loop here because the reward signal would be slow (we would only know months later whether an Intake Report caught the right thing) and noisy (engagement outcomes depend on far more than the Intake Report). Supervised tuning against senior reviewer Reports is the cleaner signal.
Performance bands
On the held out evaluation set (250 historical briefs the agent has not seen, with Reports written by senior engagement leads), the Intake Agent matches the major gap or risk identified by the human reviewer in roughly 80 to 88 percent of cases. The misses cluster in three areas: briefs in industry domains with vocabulary the agent has not seen, briefs where the gap is non textual (a missing diagram, an absent attached document), and briefs where the human reviewer drew on context outside the brief itself (knowledge about the customer’s prior engagement, awareness of an industry event the brief implicitly references).
False positives (gaps the agent surfaced that the senior reviewer judged not material) sit in the 12 to 18 percent range. The Technical Review Team treats every flagged item as something to consider, not as something to act on; the false positive rate is the cost of a sensitive reader, and we would rather the agent flag too much than too little at the front of the funnel.
The agent’s Report is generated in 2 to 4 seconds for typical brief lengths and up to 8 seconds for very long briefs (rare). This puts it well inside the latency budget where the Console can show the Report immediately after the human reviewer opens the brief.
Limits
Three structural limits.
Domain depth. When the brief leans heavily on jargon from a domain the agent has not seen (specialised manufacturing processes, niche regulatory frames, exotic technical stacks), the agent’s Report becomes less reliable. The Domain notes section is meant to surface this; the human reviewer treats the rest of the Report with appropriate skepticism in those cases.
Non textual content. The agent reads text. Briefs that include diagrams, screenshots, attached design documents or video walkthroughs are partially read. The Console captures the attachments and surfaces them to the human reviewer separately, but the agent’s Report does not incorporate them.
Implicit context. When the brief assumes context from a prior engagement, an industry event or an unstated relationship, the agent has no way to surface that. The human reviewer brings the implicit context.
Binding versus exploratory
The Intake Report is labelled exploratory in the Console. This is structural, not cosmetic. The label is what tells the customer (and the platform’s downstream layers) that the Report is the agent’s reading of the brief, not the platform’s response to it.
Binding output starts at the Technical Review Team. When the team reviews the brief and the Intake Report and decides the engagement is ready to move forward, that decision creates the first binding artifact: an Engagement Acceptance Note. The Note is signed by a named senior reviewer; the customer responds to the Note, not to the Intake Report.
This separation is the same separation we apply at every layer of the stack. Models propose, humans dispose. The Intake Agent is the first instance of the pattern; the QA Agent and the Planning Agent are the others.
Specifications
| Attribute | Value |
|---|---|
| Underlying model | Orzed Pulse, Intake fine tune |
| Training data | ~8,000 anonymised briefs with senior engagement lead Reports, plus retrospectives and negative examples |
| Median latency | 2 to 4 seconds (typical brief) |
| Output | Structured Intake Report (Summary, Gaps, Ambiguities, Risks, Domain notes) |
| Console surface | Intake Report drawer on every submitted brief, visible to the Technical Review Team |
| Binding status | Exploratory; binding decisions sit with the Technical Review Team |
The Planning Agent write up explains the next step in the engagement lifecycle, where the brief plus the Intake Report plus the Technical Review Team’s decisions are converted into a Planning Recommendation by Horizon.
Questions teams ask
Does the Intake Agent reply to me?
No. The Intake Report is for the Technical Review Team, not for the customer. The team reads the Report alongside the original brief and decides what to surface back to the customer (questions to clarify, risks to flag, scope to renegotiate). This separation prevents an exploratory model output from being mistaken for a binding response.
What kinds of gaps does the agent catch?
Missing acceptance criteria, ambiguous scope boundaries, undefined success metrics, mention of compliance domains without a stated regulatory frame, integrations with external systems whose contracts are not described, and timelines that imply throughput beyond reasonable bands. It also flags briefs that lean heavily on an industry vocabulary the agent has not seen, so the human reviewer knows to verify domain assumptions.
Is my brief used to train the model?
No. Customer briefs are partitioned and never enter the Intake Agent's training set. Fine tunes draw from anonymised historical briefs that are part of the Orzed delivery memory and have been explicitly cleared for training use, plus annotations from senior engagement leads.