Agentic AI|September 28, 2026|12 min read

Multi-Agent Color Team Reviews: AI Pink, Red, and Gold Passes

How to build three separately prompted AI reviewer agents bound to Section L, Section M, and the SOW so every Pink, Red, and Gold finding cites solicitation text instead of reviewer taste.

Rachel Kim|Director of Government Solutions

It is Thursday afternoon. The Red Team review is Friday at 9am. Two of your three senior reviewers just dropped, and the one who shows up spends ninety minutes rewriting the transition approach into her own voice without once opening Section M. You end the session with a redlined document and no idea whether the proposal would score well against the stated evaluation factors.

A multi-agent color team review fixes the first pass of that problem. You assign each color team pass to a separately prompted AI reviewer agent, bind each agent to a fixed evidence set (Section L instructions, Section M evaluation factors and their stated relative importance, the statement of work, Q&A amendments, and the draft volume), and require every finding to cite the instruction, factor, or paragraph behind it. Structured findings land days before humans convene. Humans still decide.

That last sentence is the boundary. Agents generate findings. Named owners accept, modify, or reject them. Nothing skips the human gate, and nothing goes into a submitted volume because a model suggested it.

Your Red Team Reviewer Is Rewriting, Not Scoring

Two constraints break color team reviews, and neither is about AI.

The first is reviewer scarcity. The people qualified to review a $40M task order response are the same people staffed on delivery, chasing the next capture, or already reviewing two other volumes that week. You get whoever is free, not whoever is right.

The second is substitution. A reviewer without a rubric in front of them defaults to personal style. FAR 15.304 requires the solicitation to state the evaluation factors and significant subfactors and their relative importance, and FAR 15.305 governs how the agency actually evaluates proposals against them [2][3]. The basis for award is written down. Most review sessions never read it aloud.

Agents do not get tired of reading Section M. That is the entire value proposition. A separately prompted reviewer agent, given the evaluation criteria as its only rubric, will score paragraph by paragraph against the factor at its stated weight without drifting into comma placement. It will also miss things a seasoned capture lead catches in ten seconds, which is why the human pass still happens, just against a pre-built finding list instead of a blank page.

Three Agents, Three Mandates, Zero Overlap

The failure mode with multi-agent reviews is not hallucination. It is overlap. If all three agents read all three documents and produce all three kinds of findings, you get 180 findings, half of them duplicates, severity counts inflated by style commentary, and a real Tier 1 compliance gap buried at row 94.

Give each agent one mandate and explicitly forbid the other two.

Pink agent: instruction-by-instruction tracing of every shall, must, and will in the instructions to offerors. Compliance matrix coverage gaps. Page counts, font and margin limits, file naming, required forms and attachments. It answers one question: is anything asked for missing or misplaced?

Red agent: adversarial scoring against each evaluation factor and subfactor at its stated relative importance [2]. Unsupported claims. Risks an evaluator could document as a weakness. It answers: how would a hostile but fair evaluator score this section?

Gold agent: executive readability, theme consistency across volumes, cross-volume number and name mismatches, production and submission checks including portal format requirements and accessibility of delivered artifacts where the solicitation requires conformance [12].

AgentBound evidence setProducesOut of scopeHuman owner
PinkSection L, compliance matrix, draft volumeMissing instructions, coverage gaps, format and page violations, absent formsScoring quality, prose style, win themesProposal manager
RedSection M factors and weights, SOW, draft volumeUnsupported claims, weakness candidates, factor-level score rationaleFormat checks, typos, portal mechanicsCapture lead
GoldAll volumes, style guide, submission instructionsCross-volume mismatches, theme drift, readability, production defectsNew compliance findings, rescoringVolume owners
Optional pricing agentSection B, pricing instructions, cost narrativeBasis-of-estimate gaps, math and narrative mismatchesTechnical scoring, formatPricing lead

Add the fourth agent only when its evidence set is genuinely separate. A pricing agent that reads the same technical volume as the Red agent just produces duplicates with a different header.

Build the Agent Brief From Primary Documents Only

Load the solicitation, not your knowledge base. That is the single most important prompt constraint.

Each agent gets the instructions to offerors, the evaluation factors with their stated relative importance, the work statement, every Q&A amendment, and the draft volume under review. Nothing else. No prior proposals, no industry best practice documents, no "typical DoD expectations." Then the instruction that makes the output usable: when the solicitation text does not support a judgment, flag it as unsupported rather than inventing a standard.

That one line eliminates most of the noise. An agent told to evaluate against general excellence will produce a hundred opinions. An agent told to cite Section L paragraph or Section M subfactor for every finding, and to mark anything it cannot cite, produces a shorter list you can actually work through.

Key the compliance matrix to the solicitation's own instruction text first, with FAR citations as supporting references only. Acquisition.gov is publishing FAR overhaul material alongside the regulation itself, which means authoritative text can move, consolidate, or shift into non-regulatory guidance [9][8]. A matrix keyed to section numbers goes stale silently. A matrix keyed to "Section L.3.2(b), technical approach page limit" survives renumbering.

Add a pre-submission check that every FAR citation quoted in your narrative still reads as quoted in current Acquisition.gov text, and retain a dated copy of what you relied on for each submission [8][9]. When a solicitation cites language inconsistent with current FAR text, raise it as a written question during the Q&A window instead of resolving the ambiguity yourself.

The Pink agent should consume the compliance matrix, not rebuild it. If your matrix already exists as structured requirement rows with owners and status, the agent is checking coverage against a known baseline. If it does not exist, start there. Automated compliance matrix generation turns the solicitation into traceable requirement rows, and Section M decomposition at machine speed covers how to break evaluation factors into scoreable subfactor units before any agent reads a draft.

The Finding Schema That Makes Reviews Auditable

Prose review comments cannot be sorted, counted, or dispositioned. A fixed schema can.

Require every finding, from every agent, in the same six fields:

  1. Location: volume, section, page, paragraph
  2. Citation: the Section L instruction, Section M factor or subfactor, or SOW paragraph the finding rests on
  3. Severity tier: 1, 2, or 3, defined below
  4. Evidence gap: what the text asserts versus what it proves
  5. Rewrite recommendation: specific, not "strengthen this"
  6. Assigned owner: a named person, not a team

Severity has to be defined in solicitation terms, not in feelings:

  • Tier 1: Nonresponsive or noncompliant. A required item is missing, a limit is exceeded, or an instruction is unmet. Citation to Section L or the submission instructions is mandatory. Blocks the gate.
  • Tier 2: Unsupported claim against a scored factor. The text makes an assertion that an evaluator scoring under FAR 15.305 would have no basis to credit [3]. Citation to the Section M factor is mandatory. Must be dispositioned before Gold.
  • Tier 3: Clarity, consistency, or production. Citation optional. Batched and resolved by the volume owner.

Two findings written in the schema, copy this format

Pink, Tier 1. Location: Vol II, Technical, Section 3.4, p. 22. Citation: Section L.4.2(c), "offeror shall provide a staffing matrix identifying labor category, clearance level, and FTE by task area." Evidence gap: narrative describes staffing approach; no matrix provided in Vol II or appendices. Rewrite: insert Table 3-4 with the three required columns keyed to SOW task areas 4.1 through 4.6. Owner: J. Alvarez. Red, Tier 2. Location: Vol II, Section 2.1, p. 9. Citation: Section M, Factor 2, Subfactor 2b, Transition Risk Management (stated as more important than Subfactor 2c). Evidence gap: claims "proven 30-day transition" with no reference to a past performance citation, no phased schedule, and no named transition lead. An evaluator has nothing to credit and could document this as a weakness. Rewrite: tie the claim to PP Reference 2 (same agency, comparable scope), add the 30-day phase table, name the transition lead with role duration. Owner: M. Chen.

Structured findings roll into one disposition log per volume. Every row carries an accept, modify, or reject decision with a date and a name. That log is the artifact you keep, and it is the thing that lets you reconstruct how a decision was made six months later.

Calibrate on a Closed Bid Before You Trust a Live One

Do not put agents on a live pursuit first. Run them against a bid you already submitted where you still have the human color team notes and, ideally, the debriefing record.

FAR 15.506 governs postaward debriefings and defines what an unsuccessful offeror can request and when [11]. If you requested one and documented it, you have the closest thing to ground truth available: what the evaluator actually wrote down as a weakness or deficiency. That is your answer key.

Score agreement three ways:

  • Both caught it. Your prompts are working for that finding class.
  • Only humans caught it. Prompt gap. Usually means the evidence set was incomplete or the mandate was too narrow.
  • Only the agent caught it. Verify before crediting. Some of these are genuine finds that a tired reviewer missed. Some are fabrications, and you need to know your fabrication rate before you trust the output.

Then tune severity definitions until Tier 1 agent findings line up with what the evaluation record treated as compliance failure. If the agent calls six things Tier 1 and the debriefing documented one deficiency, your Tier 1 definition is too loose and reviewers will start ignoring the tier entirely.

Calibration scenario. A team ran the three agents against a closed IT services task order bid. The Red agent flagged Factor 3, Subfactor 3a (Relevant Corporate Experience) because the narrative claimed comparable scope at a named agency but cited no past performance reference, and the two references submitted covered a different task area. The human Red team had marked the same section "needs more detail" with no citation. The debriefing had documented exactly that subfactor as a weakness. That single agreement told the team their Red prompt was aimed correctly and their human notes were the weaker artifact.

Load debriefing feedback into a lessons database keyed to evaluation factors and feed it to the Red agent as standing calibration input on the next pursuit [11]. Past performance evidence is where this pays off fastest, and CPARS-aligned past performance automation covers how to keep those references retrievable instead of reconstructed each cycle.

Where the Human Gate Stays Non-Negotiable

Assign a named owner per volume. Not a team, not a distribution list. One person who accepts, modifies, or rejects every finding and records the disposition with a date.

Route unresolved Red team disagreements to the capture lead, never to another AI pass. Running a second agent to adjudicate the first one is how you end up with a confident answer nobody can defend. A human with authority makes the call and the log records who made it.

Verify every factual claim, quantity, certification, and past performance reference by hand. Fabricated citations and cross-volume number mismatches are the failure modes that cost evaluations, and no agent mandate reliably catches its own invention. If a number appears in the technical volume, the management volume, and the price narrative, someone checks all three.

Keep accountability at the artifact level: named human author and named reviewer per section, plus the retained color team disposition log. When an agency asks who is responsible for the content in a submitted volume, the answer is a person's name, not a tool.

Governing the Agents When the Agency Asks How You Use AI

Agencies are starting to ask contractors how they govern generative AI on proposal and delivery work. Teams that improvise on that call sound worse than their process actually is.

Publish an internal AI use policy before the first pursuit. It needs four things: approved tools, prohibited data categories, required human review before AI-influenced text enters a submitted volume, and named ownership of the policy itself. Make the prohibited categories explicit rather than implied: customer-sensitive material, export-controlled technical data, personally identifiable information, and competitor-provided information.

Map your practice to the NIST AI Risk Management Framework functions and the AI 600-1 Generative AI Profile so your answer uses shared vocabulary instead of a framework you invented [4][5]. When a contracting officer asks how you manage AI risk, describing your review gates in govern, map, measure, and manage terms is faster and more credible than a narrative.

Read current federal AI guidance, including GSA's published AI resources, to understand what your customers are being told to ask [7]. Then describe your process in that language without asserting requirements the guidance does not impose. Overclaiming compliance with a policy that does not apply to you is its own risk.

Set retention and access rules up front: who may run the agents, what material enters which tool, and how long outputs are kept. And keep the traceability discipline, because if an award is later challenged, the record has to speak for itself. GAO's bid protest process turns on the documented record, not on remembered intent [6]. Solicitation-cited findings with dated dispositions are a better record than a marked-up Word file. Our own position on AI accountability in proposal work is documented on the trustworthy AI page.

Frequently Asked Questions

Can AI agents replace human color teams?

No. They replace the first pass of finding generation. Decision authority, factual verification, and disagreement resolution stay with named humans. Documented roles, inputs, and human accountability for AI-assisted output are what the NIST AI Risk Management Framework and its Generative AI Profile ask an organization to establish [4][5].

How many agents do you need?

Three mandates is the working minimum: Pink for compliance tracing, Red for adversarial factor scoring, Gold for production and cross-volume consistency. A fourth pricing or teaming agent helps only when its evidence set is genuinely separate from the other three.

What if the solicitation is ambiguous?

The agent flags it as unsupported and you raise a written question to the contracting officer during the Q&A window. FAR Part 15 covers exchanges with industry in a negotiated procurement, including exchanges before proposals are received [1]. Do not resolve solicitation ambiguity internally and hope the evaluator agrees.

How do you stop agents from duplicating each other?

Non-overlapping mandates plus explicit out-of-scope instructions. A Pink agent told "do not comment on scoring quality or prose style" stays in its lane.

What evidence does the Pink agent need before it can run?

Structured requirement rows, not a draft volume alone. The Pink agent checks coverage against a known baseline, so it needs a compliance matrix keyed to the solicitation's own instruction text with an owner and a status on every row. If that baseline does not exist yet, automated compliance matrix generation is the prerequisite step, because a coverage gap can only be reported against a requirement someone already recorded.

What does this change about the human review meeting?

Reviewers arrive with a sorted finding list and spend the session dispositioning Tier 1 and Tier 2 items instead of reading cold. Every disposition carries a name and a date, which is the record that has to speak for itself if an award is later challenged [6].

Your First Calibration Run This Week

Pick one closed bid where you still have the color team notes. Write the three role prompts. Run all three agents against the submitted volume. Build the three-column agreement table: both caught, humans only, agent only. That is a half day of work and it tells you more about your review process than your last four Red Teams combined.

Track one metric going forward: the percentage of Tier 1 findings closed before the human Pink team convenes. If that number is climbing, your agents are doing the work reviewer scarcity used to prevent. If it is flat, your evidence sets are incomplete or your severity definitions are not tied to solicitation text.

Then fix the input. Findings are only as good as the requirement baseline behind them, so make sure your matrix rows and draft structure are built to be checked. Structured drafting keeps sections mapped to the instruction and factor they answer, which is the condition that makes every agent finding citable in the first place.

The Friday review will still happen. It will just start with a list instead of silence.

References