What is an AI subagent?

An AI subagent is a second AI worker that your primary agent hands a complete job to. It works in its own separate space with its own toolset, does the heavy lifting without cluttering your conversation, and returns one finished report. The pattern mirrors how a capture manager hands a research assignment to an analyst: a scoped task, the access to do it, a defined deliverable, and a report you can review.

How is a subagent different from asking a general-purpose chat assistant a research question?

The comparison is really about connection and structure, not about which model is smarter. An unconnected chat session knows only what you paste into it. A subagent works against real procurement data (contract records, award history, contractor and agency profiles, government contacts, and the actual solicitation attachments), so it is researching rather than recalling. It returns a structured report with claims traced to sources, including a gaps section when the data could not answer something. And it works in its own context, so a long investigation does not degrade the quality of your main conversation the way one long chat thread does. If you prefer to work inside ChatGPT, the CLEATUS ChatGPT app connects that same procurement data to your session.

How many subagents can run at once?

CLEATUS runs up to three subagents at the same time on independent tasks. That is why a question like "compare the incumbent, the agency buying history, and the solicitation requirements" comes back as one answer rather than three sequential waits.

Can I see what a subagent is doing while it works?

Yes. Each subagent appears as a single live line that updates with its current activity, plus a timer. Click it to open the full run: the exact task it was given, its reasoning, and every search it made. You can stop a run at any point and it ends cleanly. That visibility is what makes the output auditable rather than a black box, which matters given that only about a quarter of contractors report mature AI governance structures.

What work should stay with humans?

Judgment and accountability. Bid and no-bid decisions, pricing strategy, teaming commitments, customer relationships, and anything you sign. Subagents are good at bounded research: incumbent analysis, award history, agency and contact mapping, solicitation requirement extraction, and teaming candidate shortlists. The value is that senior people get those hours back for the decisions only they can make.

When should I use a workflow instead of a subagent?

Use a subagent when the job would need a different explanation every time, such as a deep dive on one specific pursuit. Use a workflow when you would explain it identically every time, such as screening every new recommendation against your bid criteria or watching a solicitation for amendments. Subagents run in your chat while you watch. Workflows run unattended on a trigger.

Does delegating research to AI create a compliance risk?

The risk comes from unverifiable output, not from delegation itself. A report whose claims trace to specific contract numbers, award values, dates, and document sections can be checked before it informs a decision. A summary with no sources cannot. Write down which decisions stay with humans, require traceable evidence for anything that feeds a bid decision, and keep the run transcripts. That is most of a workable AI governance posture for a capture team.

AI Agents for Government Contractors: How…

Q: What does it mean that a solicitation is "pre-shredded"?

It means the reading already happened. When an opportunity matches your company, CLEATUS automatically scans every page of the package (the base solicitation, the amendments, the attachments, and the exhibits, including pages that are scanned images rather than text) and breaks it into one standardized Contract Breakdown following the Uniform Contract Format: scope, requirements, deliverables, deadlines, pricing, and evaluation criteria. Your fit score is computed against that breakdown, so it reads Sections C, H, K, L, and M plus the attachments instead of a title and a NAICS code. Most platforms shred only when a user triggers it on one opportunity, which means the analysis exists only for pursuits somebody already chose to chase. Reading the package is how you decide whether to chase it, so the order matters.

The 2026 Deltek Clarity GovCon Study, based on 917 contractors surveyed in January, describes a bind that every capture lead will recognize. Revenue grew about 15% in 2025 and firms expect 16% in 2026, while average profit margins fell from 20% to 17%. Nine out of ten firms saw at least one financial metric decline. Contractors reported an average of 84 hours to develop a single proposal, and 83% said they missed opportunities in 2025 because they found them too late. More work is arriving, each pursuit costs more to chase, and there is less margin to hire against. When you cannot add headcount and you cannot do less work, the only remaining lever is delegation. That is the real problem AI agents and subagents are built to solve, and the reason a chat subscription on its own has not moved anyone's win rate.

TL;DR

GovCon is structurally a delegation business. Capture, proposal management, pricing, contracts, and compliance are separate disciplines because no one person can hold all of them. In a large prime, three teams do that work. In a small business, one person does all three, which is normal and is also the growth ceiling.
The 2026 numbers make delegation the only available lever. Margins compressed from 20% to 17% while revenue grew 15%, so the classic answer of hiring another analyst costs more than it used to and returns less.
AI adoption is not the same as delegation. 90% of contractors now use AI in some capacity, up from 45% a year earlier, but only 5% call their AI capabilities fully developed and 45% say they are unclear on the return. Firms bought a tool. They did not hand over a job.
Real delegation has four conditions: a scoped task, the access needed to do it, a defined deliverable, and a way to check the work. An unconnected chat thread fails all four, which is why prompting feels productive and rarely compounds.
AI subagents are delegation with those conditions built in. The GovCon Copilot hands a research-grade question to a focused subagent that works in its own space with its own research tools, runs up to three at a time on independent tasks, and returns one complete report instead of a pile of raw search results.
The package is read before you delegate. Every matched opportunity arrives pre-shredded: CLEATUS scans every page of the base solicitation, the amendments, the attachments, and the exhibits, including scanned pages that are images rather than text, and returns one standardized Contract Breakdown. Competitors shred when a user triggers it, one pursuit at a time. Automatic beats user-triggered, because triage is where you cannot afford to have not read yet.
You keep the audit trail. Each subagent appears as a live line you can open mid-run to see the exact task it was given and every search it made, and you can stop it. That matters when only a quarter of contractors report mature AI governance.

Delegate a Real Research Job

Watch CLEATUS split one capture question into parallel subagents, work the incumbent, the agency, and the solicitation at the same time, and hand back one answer you can trace.

Book a Demo →

GovCon Has Always Been a Delegation Business

Look at how the industry organizes itself and you are looking at a delegation structure. A capture manager owns the pursuit from qualification through proposal kickoff: competitive analysis, stakeholder mapping, teaming strategy, gate reviews. A proposal manager owns the document: outline, compliance matrix, writer assignments, color team reviews, production. Pricing owns the cost volume. Contracts owns the representations and certifications. Compliance owns the clause flowdowns. These are separate roles because the cognitive load of a federal pursuit does not fit in one head, and the industry figured that out decades ago.

The trouble is that the structure assumes people to fill it. In a large prime, three different teams handle capture, proposal management, and pricing. In a small or mid-sized business, one person does all three, and that is completely normal. The Association of Proposal Management Professionals frameworks that describe clean role boundaries are written for an org chart most contractors do not have.

So the small firm improvises. The owner is the capture manager. The capture manager is also the proposal manager. The proposal manager is also the person who reads the PWS appendix at 11 PM to find out whether a facility clearance is required. Every hour that person spends on research is an hour not spent on the two things nobody else can do: the bid decision and the customer relationship.

This is the delegation ceiling, and it is the same thing as the growth ceiling. A firm can pursue exactly as many opportunities as its senior people can personally think about. Add a fourth simultaneous pursuit and something gets a shallow read. Add a sixth and the shallow reads start producing bad go decisions, which is the expensive kind of mistake: a full proposal cycle spent on work you were never going to win.

The Math That Makes Delegation the Only Lever Left

For most of the last decade, the answer to the delegation ceiling was to hire. Bring on a junior capture analyst, hand them the research, and buy back senior hours. The 2026 Deltek Clarity data explains why that answer is getting harder to afford.

2026 Deltek Clarity finding	What it means for delegation
Average profit margin fell from 20% to 17%	Every new hire consumes a larger share of a thinner margin, so the payback period on a junior analyst stretches
Revenue grew about 15% in 2025, 16% expected in 2026	Pursuit volume is rising faster than teams are growing, which widens the gap the senior people absorb
84 hours average to develop one proposal	Two full work weeks per bid, and a large share of it is research and assembly rather than persuasion
83% missed opportunities in 2025 due to late discovery	The work that gets dropped first is the upstream research nobody has bandwidth to delegate
9 in 10 firms saw at least one financial metric decline	Capacity has to come from somewhere other than the payroll line

Read those together and the conclusion is uncomfortable but clear. Firms are getting bigger without getting more profitable, as Deltek's Kevin Plexico put it. Pursuit volume is climbing, the cost of each pursuit is climbing, and the margin available to hire against is falling. You cannot solve that by working later. You solve it by moving work off the senior people who are the constraint, and the only capacity available at the price the margin allows is machine capacity.

AI Adoption Is Not the Same Thing as Delegation

Here is the part of the 2026 data that should stop a capture lead cold. AI adoption roughly doubled in a year, from 45% to 90% of contractors using AI in some capacity, and more than 90% report using generative AI. Yet only 5% describe their AI capabilities as fully developed, only about a quarter report mature governance structures, and 45% say they are unclear on the return.

That is not a technology failure. It is a delegation failure. Buying a chat tool and telling the team to use it is the organizational equivalent of hiring a contractor, giving them no brief, no system access, no deliverable format, and no review, and then wondering in the quarterly why nothing changed.

When people say AI did not work for their capture team, what usually happened is that nothing was ever actually handed over. Someone pasted a section of a solicitation into a chat window, got a summary, decided it was roughly right, and did the real work themselves anyway. That is not delegation. That is a second opinion with extra steps. It shows up as a tool subscription and not as a single reclaimed hour, which is exactly the ROI ambiguity 45% of contractors report.

What Good Delegation Actually Requires

Before talking about software, it is worth writing down what a competent capture manager does when handing a job to a junior analyst. The pattern is consistent across every firm that delegates well, and it has four parts.

A scoped task with a stated purpose. Not "look into NAVFAC," but "find out who has held NAVFAC Mid-Atlantic janitorial work for the last five years, what those awards were worth, and whether any of them are small business set-asides, because I need to decide whether we bid the recompete." The analyst needs the decision the work feeds, or they optimize for the wrong thing.

The access to do the work. A junior analyst who cannot get into SAM.gov, cannot see your award history, and cannot open the solicitation attachments will hand back something confident and wrong. Access is not a convenience. It is what separates research from guessing.

A defined deliverable. A memo with the incumbent named, the award values listed, the set-aside status stated, and gaps flagged where the data did not exist. Analysts who are told "just tell me what you find" produce whatever length feels appropriate, which is never what the reviewer needed.

A way to check the work. Every claim traceable to a source. Not because the analyst is untrustworthy, but because you are the one who signs the bid decision, and a capture lead who cannot verify a finding cannot responsibly act on it.

Skip any one of these and delegation degrades into supervision, which costs more than doing the job yourself. That is true for a human analyst and it is exactly as true for an AI one. It is also a fair test to run against any AI tool a vendor puts in front of you.

Why a Single Chat Thread Cannot Be Delegated To

Hold an unconnected, general-purpose chat session up against those four conditions and it fails all four, for reasons that have nothing to do with the quality of the model.

It has no access. A general-purpose chat session, opened cold, cannot see your award history, your past performance library, your capture profile, or the amendment posted to a solicitation this morning. It knows only what you paste into it, and real solicitations are multi-document packages where the disqualifier usually sits in an attachment or an amendment rather than the synopsis. Access is the variable here, not the model. Connect that same assistant to real procurement data and the picture changes, which is exactly what the CLEATUS ChatGPT app does.

But connection alone is only half of access, and the missing half is timing. Being able to fetch a document when someone finally asks is not the same as having already read every page of the package before anyone asked. That difference decides what a delegate can find, and it is worth its own section.

It has no defined deliverable. Ask the same research question twice and you get two different shapes of answer, which means a human has to normalize the output before it can be used in a gate review. That normalization is the work you were trying to delegate.

It gives you nothing to check. A capture memo with no traceable sources is a memo you have to re-research before you can put your name on it.

And the fourth failure is the one people notice last: a single thread degrades as the work gets big. Every search result, every document excerpt, and every dead end piles into the same conversation. Somewhere in hour two the answers get vaguer, earlier findings stop being referenced, and the thread that was sharp at the start starts contradicting itself. This is the practical reason long investigations in one chat window feel great for twenty minutes and unreliable by the end. It is also why you cannot run three investigations at once in one thread: they contaminate each other.

A human org chart solved this problem a long time ago by giving each person their own desk, their own assignment, and their own file. The fix for AI is the same shape.

AI Subagents: Delegation With the Structure Built In

A subagent is a second AI worker that your primary agent hands a complete job to. It works in its own separate space with its own toolset, grinds through the job without any of that traffic reaching your conversation, and comes back with one finished report.

In CLEATUS, when you ask the GovCon Copilot a research-grade question, it tells you it is delegating and spins up focused subagents, running up to three at the same time on independent tasks. Map that against the four conditions:

Scoped task. The Copilot splits your question into separate assignments and writes each one out explicitly. You can see the exact task each subagent was given, which means you can also tell when it split the work wrong and correct it in one sentence.

Access. Each subagent carries a research toolset against real procurement data: contract search and full contract detail, award history and semantic award search, contractor profiles and award records, agency profiles, government contact and contracting officer lookups, NAICS lookup, document outlining, document reading, keyword and semantic search inside solicitation attachments, plus web search for anything outside the platform. It reads the actual package rather than the paragraph you remembered to paste, and it does not start from raw files, because the package was already scanned and segmented when the opportunity matched you. Everything a subagent can reach is scoped to your organization the same way the rest of your workspace is, and the tools that take real action stay with the primary agent rather than being handed to a delegate.

Defined deliverable. A subagent returns a structured report: the direct answer first, then sections per question with the supporting evidence inline, award values, dates, NAICS codes, contract numbers, document names and sections. Where the available data could not answer something, it says so under a gaps section rather than filling the hole with a plausible sentence. Your agent then folds the reports into one answer, with charts where they help.

A way to check the work. Each subagent shows up as a single live line that updates with what it is doing right now. Click it and you see the full run: the task it was given, its reasoning, and every search it made. Change your mind mid-run and you can stop it cleanly.

The context benefit is the quiet one, and it is the reason this works at all. Because the heavy searching happens in the subagent's own space, none of it crowds your conversation. Your agent stays sharp through a long back and forth, and you can have three investigations running side by side without them bleeding into each other. Work that used to mean an afternoon of tab hopping across SAM.gov, USAspending, and a half dozen agency pages comes back as one brief while you keep working.

The Package Is Already Read Before You Delegate Anything

Here is the part that decides whether any of this works, and it happens before you ask your first question.

A delegate can only find what it can read. Hand a human analyst a solicitation and the first day disappears into opening files: the base document, the amendments, the attachments, the exhibits, the wage determinations, the drawings. Some of those arrive as scanned images rather than text, so searching them returns nothing at all and somebody has to read them with their eyes. This is the single most common way a disqualifier survives to bid day. The requirement was in the package. Nobody got to page 240 of an exhibit.

CLEATUS reads the entire package on the match, automatically, before anyone opens the opportunity. Every page of the base solicitation, every amendment, every attachment, and every exhibit gets scanned, including the pages that are images rather than text, and the result is broken into one standardized Contract Breakdown that follows the Uniform Contract Format: scope, requirements, deliverables, deadlines, pricing, and evaluation criteria. This is what "pre-shredded" means. The work is finished before the opportunity reaches you.

That changes three things that matter for delegation:

Your fit score is computed against the actual package. The breakdown is what the score reads, which means Sections C, H, K, L, and M plus the attachments, rather than a title, a NAICS code, and a set-aside. A score derived from a cover sheet is a guess dressed up as a number.

A subagent starts from an analyzed package, not a stack of files. When you delegate a solicitation deep dive, the subagent is not opening PDFs for the first time and hoping the text layer is readable. It is searching a structure that already exists, which is why a question like "what in the attachments would disqualify us" returns something in minutes instead of being the reason nobody asks.

When the package has not been read, the card says so. Confidence drops rather than the model filling the gap with a plausible sentence. That is the same discipline as the gaps section in a subagent report, applied one layer earlier, and it is the difference between a system you can delegate to and one you have to double-check.

The contrast with the rest of the category is not subtle. Most platforms that advertise shredding run it when a user triggers it, one opportunity at a time, usually as the first step of starting a proposal. That means the analysis exists only for the pursuits somebody already decided to chase, which is backwards: the reason to read the package is to decide whether to chase it. Automatic beats user-triggered because triage is exactly where you cannot afford to have not read yet.

Four Jobs Worth Handing to a Subagent

Delegation pays off most on jobs that are bounded, research heavy, and low on judgment. These four qualify, and they map to the work that most often gets skipped when a capture lead is running four pursuits at once.

1. The solicitation deep dive. "Research this solicitation thoroughly: what are the mandatory qualifications, what is the evaluation scheme, and what in the attachments would disqualify us?" Because the package was already scanned and segmented into a Contract Breakdown when the opportunity matched you, the subagent is reading a structure rather than opening files. It works the FAR section breakdown, the amendments, and the attachments, and searches inside them for requirements, evaluation criteria, and compliance language. The buried disqualifier is the whole point of this one.

2. The incumbent and competitor workup. "Who currently holds this work, what else have they won at this agency, and how big are they?" The subagent finds the firm, pulls its registration data and award history, and compares that history against the pursuit you are considering. This is the research that most reliably changes a bid decision and most reliably does not get done.

3. The agency and contact map. "Map this buying organization: which offices buy what we sell, who are the contracting officers and program staff, and what have they awarded in our codes over the last three years?" Useful in its own right, and much more useful 12 to 18 months before a recompete, which is where procurement forecasting actually pays.

4. The teaming shortlist. "Find firms with past performance in this scope and the right set-aside status that we could team with on this pursuit." Teaming research is high effort, spread across several data sources, and almost always compressed into the last week before submission. It is a strong delegation candidate for exactly that reason. Our guide on how to find teaming partners in government contracting covers what to do with the shortlist once you have it.

A practical note on prompting: say what decision the research feeds. "Research the incumbent" produces a profile. "Research the incumbent because I need to decide whether to bid the recompete or approach them about a subcontract" produces an argument. The rule is the same one you would apply to a human analyst, which is the point.

Subagents and Workflows: Delegating With You, and Without You

Subagents are one of two delegation modes, and choosing the wrong one is a common source of disappointment. Subagents run inside your chat while you watch and steer. CLEATUS Workflows run unattended, started by an event rather than by you, and the full write-up on Workflows covers what teams build with them.

	Subagents	Workflows
Who starts it	You, in conversation	An event: a new recommendation, a posted amendment, a saved-search match, a pipeline change, a schedule
Are you present	Yes, watching and able to stop it	No, it runs with nobody there
Shape of the job	Open ended research on one pursuit or question	A repeatable process you have defined once
Best used for	Deep dives, incumbent workups, agency maps, teaming shortlists	Intake triage, go/no-go screening, amendment monitoring, pipeline reporting

The rule of thumb: if you would have to explain the job differently every time, delegate it to a subagent. If you would explain it the same way every time, encode it as a workflow. Most teams end up running both, and the real workflow examples our customers built are a reasonable place to see what belongs in the second column.

What Delegation Looks Like When It Works

The firms that have moved furthest on this are not the ones with the largest teams. They are the ones that stopped treating research as senior work.

D2 Government Solutions tripled proposal output with the same team, cut discovery time by 75%, and cut draft time by 80%. MST Maritime Management went from three proposals per month to more than ten, a 4× increase, with the same lean team. Operation Hired reached 6× proposal throughput. In every case the headcount stayed flat and the research load moved.

"Before CLEATUS, we were spending almost our entire week just hunting for opportunities and trying to understand what each solicitation was asking for. All that upfront work left us with very little time for actual proposal development. We were lucky to complete three proposals per month."

– Miguel Morgan, CEO, MST Maritime Management

That quote is a delegation diagnosis, not a software complaint. An entire week consumed by hunting and understanding is a week of senior time spent on work a well-briefed analyst could have done. The capacity was never missing. It was allocated to the wrong task.

A caution worth stating plainly: delegation moves the work, not the accountability. The bid decision, the price, the teaming call, and the signature stay with you. A subagent that produces a clean incumbent analysis has not made your go/no-go decision, and any tool that claims otherwise is selling you a risk you will personally own at debrief.

How to Start Delegating Without Losing Control

Given that 73% of contractors are still early on AI governance, the sequencing here matters more than the enthusiasm.

Pick one recurring research job, not a category. The incumbent workup is a good first choice: bounded, high value, and easy to grade because you often already know the answer for one or two pursuits.
Delegate it on a pursuit you have already researched by hand. Compare the subagent report against what your team produced. You are calibrating trust with a known answer, which is the same thing you would do with a new analyst.
Open the run. Read the task it was given and the searches it made. If the answer is wrong, the transcript usually shows exactly where: a bad scope, a missing document, a search that never happened.
Fix the brief before blaming the tool. Most weak reports trace back to a task that did not state the decision it fed. Rewrite the request the way you would rewrite an assignment to a junior analyst and run it again.
Write down what you will not delegate. Bid decisions, pricing strategy, teaming commitments, and anything a customer relationship depends on. A short written boundary is most of what an AI governance policy needs to say, and it is far more useful than a policy nobody reads.
Promote the repeatable ones into workflows. Once you have delegated the same job three times with the same brief, it is a process, and it should be running on a trigger rather than waiting for you to ask.

The firms that will pull ahead over the next two years are not the ones with better prompts. As we argued in Stop Prompt Engineering. Start Winning Contracts., the leverage was never in the phrasing. It is in how much of the pursuit you can responsibly hand off, and in whether you can check the work when it comes back.

Frequently Asked Questions

Government contracting was always going to reward the firms that delegate well. What changed in 2026 is that margins stopped funding the traditional way of doing it. The teams pulling ahead are handing off the research, keeping the judgment, and checking the work every time.

Book a Demo →

Customer Stories

About CLEATUS

CLEATUS is an agentic AI platform that helps government contractors discover the right opportunities, manage capture pipelines, and write winning proposals. Automate your GovCon operations end-to-end: build custom AI-powered automations that handle multi-step processes on autopilot, a force multiplier for your capture and BD team that works in the background while you stay in the loop on every decision. We aggregate federal, state, local, and city opportunities; our GovCon Copilot analyzes solicitations and your internal documents to deliver actionable market intelligence that drives revenue growth.

AI Agents for Government Contractors: How Subagents Work