Strategy AI agent · deep research
Deep Research AI Agent
Fetch what each footnote points at before a claim enters the brief, and file the claim beside the document itself, the day it was retrieved and whether that document is primary.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Researcher-held
- Retrieval-first
- After launch
- Agent Care
What this agent does
Opens what is cited, never rules on the claim
In
- 01
A claim arrives in a draft, and the citation under it is treated as an object to be fetched.
- 02
A footnote is offered, and it is opened, read and marked with the day it was retrieved.
Reason
- 03
A summary of a study is to hand, and the study itself is marked as still not held.
- 04
A sentence is quoted, and the words in the brief match the page or the line is marked paraphrase.
- 05
A document sits behind an entitlement, and the brief says where it lives and who may reach it.
Decide
- 06
A position taken in another country is described, and the jurisdiction it belongs to travels on.
- 07
A text is read in translation, and the original language and the translating hand are both named.
Out
- 08
A search returns nothing, and where the agent looked is written down as the finding it is.
- 09
Execute write actions only inside the approval boundaries agreed during implementation.
Product statement
Retrieval, labelling and drafting belong to the agent, which sets out what it searched rather than claiming to have found everything. Each claim is accepted into the brief by a named researcher.
Example workflow
One claim, citation to acceptance
- 1Citation evidence receivedPublic registers, licensed archives, prior briefs or documents named in a draft
- 2Claim context assembledThe claim, the brief it serves, the citations put forward for it and where each one now lives
- 3Sourced claim draftedThe claim, its citations, what each one resolved to and completeness
- 4Controls appliedRetrieval checks, primary-source checks, quotation checks and completeness confidence
No human action required
Stages 1 to 4 run unaided, and no claim is admitted at any of them — the agent is fetching, and a reviewer joins at the completeness gate.
Passes to a named researcher for acceptance.
Adds a research lead read first.
Research review
The claim is held with its citations, what each one opened onto and the day it was fetched.
Accept · Append citation · Send to research review
Into the brief — by a named researcher
- 6Claim and citation records updatedOnly where write access and records policy allow it
- 7Outcome evaluatedRetrieval outcomes, primary-source coverage, corrections made and what the review turned up
Corrections
Each reviewer correction is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Accepting a claim into the strategy brief.
- Deciding a secondary summary is good enough.
- Judging a source credible on its reputation.
- Signing off the brief a board will read.
Automation boundary
Agent acts unaided
- ✓Fetch every cited document that a reader could open.
- ✓Mark each source primary or secondary and say which one is in hand.
- ✓Reproduce quoted words exactly and label every paraphrase as paraphrase.
- ✓Record the retrieval date and the language read.
Nothing enters the brief and nothing is published except by a named person on approval.
- Ruling on whether a claim is true.
- Telling the board what the evidence means.
- Setting the sourcing bar a brief must clear.
- Changes to the brief or the citation record.
Example output
One claim, annotated
The market research agent we already ship serves an insights function working on commercial market figures; this one serves a strategy team, and its subject is whether a cited thing can actually be fetched and read.
Brief entry · single claim
Illustrative example- Claim
- Competitor capacity, desk research
- Recorded as
- Retrieved and read in full
- Citation
- Primary source
- Evidence of record
- Fetched 3 August 2026
- Confidence
- Held unaccepted
- Held for
- The researcher who accepts, by name
As received
Taken from the document that was opened and dated on retrieval, and nothing beyond it is implied.
What the record holds
- Primary document
- Retrieval date
- Quotation checked
Why the claim waits here
What a retrieved document means for a decision is a judgement people make.
What the score decides
Under the configured threshold the claim collects a research read before it is accepted.
Action
AcceptAppend citationSend to research reviewValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every claim
From the document behind it
Where the evidence is used
A brief is read long after the decision it informed, so each claim carries the day its source was retrieved, and the agent reports what it searched rather than claiming it found everything.
A chain ends somewhere
Walking a reference backwards stops at the document that first stated the thing, and the agent writes down where it stopped and why it could go no further.
What was checked, and not found
One translated report was never matched to an original in the issuing archive, one working paper quoted across a sector has no publisher of record, and two licensed collections could not be opened from outside a member institution.
Build an evidence trail
The claim, the primary source under it and the researcher who accepted it stay on file.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Public and published sources
- Registers
- archives
- statistical offices
- Primary documents and their dates
Licensed and subscription corpora
- Journal platforms
- news archives
- Full text behind an entitlement
Internal knowledge
- Document stores
- earlier briefs
- Prior claims and their sources
Agent
Deep research sourcing
- Reads the brief
- Fetches the sources
- Holds for acceptance
Brief and record systems
- Research library
- ticketing
- Claim and acceptance records
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six combs between the model and the brief
Six combs drawn through in order, the last the finest. Whatever remains is set out in the map below.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether a claim may stand
- L3
- Hands acceptance to a named person
- L4 – L5
- Keep the claim and the source behind it
- L6
- Returns to citation listing when signals degrade
How Nestack evaluates it
Evaluate the whole assembly — not only the claim that comes out.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the claim a brief carries
- E1Final-output evaluationDid the brief record what a later reader would need to open?
- E2Step-level evaluationDid the agent reach the primary document, or only something written about it?
- E3Tool evaluationDid the fetch succeed, and was the result stored against the right claim?
- E4Confidence calibrationDo low-confidence claims actually attract more research corrections?
- E5Slice evaluationHow does performance change across specific claim types?
- E6Business outcomeHow many claims were sent back before a researcher would accept them?
Floor — the brief a decision rests on
Failure modes
Where each failure originates in the agent
Seven failure modes, placed where each one first becomes visible.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- ME-03
Superseded edition fetched
The version opened is no longer the current one.
Stage gathers
The briefs, the claims, the citations offered
02 · Reasoning
2 modes
- ME-04
Citation resolves to nothing
A reference looks right and opens on an empty page.
- ME-06
Summary stands for document
A secondary account is used where the primary was needed.
Stage proposes
The claims, their citations and completeness
03 · Tool / write
2 modes
- ME-02
Unread citation carried on
A claim advances before its source has been opened.
- ME-05
Quotation drifts from source
Words in the brief are not the words on the page.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- ME-01
Accepted, source unopened
The brief shows a claim but nothing that was read.
Stage returns
The claim a brief carries and a reader checks
05 · Change / Version
1 mode
- ME-07
Silent retrieval regression
A rules change relaxes the fetch test without a note.
Stage tracks
Model, prompt, retrieval rules and claim fields
- Sev-1 · a claim cited to nothing at all
- Sev-2 · a paraphrase presented as a quote
- Sev-3 · source unreachable, claim held back
Affected slices
Secondary summaries absorb the corrections
A question-level retrievability figure can read clean while claims resting on secondary summaries carry most of the rework. Nestack reports the correction rate by claim type, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Claims resting on secondary summaries | 8.7% | 3.6× | Review | |
| Claims quoted from translated sources | 6.2% | 2.6× | Review | |
| Claims taken from paywalled archives | 3.9% | 1.6× | Watch | |
| Claims with a primary document in hand | 2.1% | 0.9× | Normal |
Bar: correction-rate lift vs. primary-document baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
What an unreachable source costs
The loop shuts when the unreachable citation is a regression case. That suite is what the next brief issued is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Correction rate rises on claims resting on summaries.
- 02
Diagnose
The footnote that resolves to a page that has never existed is taken apart until a single cause is left.
- 03
Improve
Any change goes out numbered, with the claims that prompted it attached.
- 04
Verify
A single failing claim case is enough to hold the release back.
- 05
Learn
One case joins the suite, one line joins the sourcing record.
Learn → Detect
The return edge. The next brief is measured against a suite one case longer.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, claim assembly, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Retrievability-standard and automation-boundary work.Week 1
- 02Public, licensed and internal sources.Week 1, Week 2
- 03Claim-to-citation and retrieval-coverage mapping.Week 1, Week 2
- 04Citation and document ingestion.Week 2
- 05Claim, citation and date binding.Week 2, Week 3
- 06Retrievability scoring and review routing.Week 3
- 07Researcher acceptance workflow.Week 3, Week 4
- 08Research-library integration.Week 5
- 09Citation and retrieval cases.Week 4, Week 5
- 10Guardrails and sourcing controls.Week 4
- 11Claim-trail instrumentation.Week 5, Week 6
- 12Deployment, documentation and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne claim type, one brief | ProductionProduction brief workflow | AdvancedMultiple briefs / languages |
|---|---|---|---|
| Introduced at Pilot | |||
| Citation retrieval to your standard | ✓ | ✓ | ✓ |
| Release by a named researcher | ✓ | ✓ | ✓ |
| Question-scope baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by research question | — | ✓ | ✓ |
| Brief acceptance workflow in your systems | — | ✓ | ✓ |
| Approved write-back | — | ✓ | ✓ |
| Retrieval-corpus integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-source briefs | — | — | ✓ |
| Cross-language evidence packs | — | — | ✓ |
| Large citation estates | — | — | ✓ |
| Multi-language sourcing controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, brief volume, approval controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01The briefs you produce and the questions behind them
→Brief inventory mapping and citation capture
Week 102Representative briefs, citations and source documents
→Citation binding, retrieval logic and the coverage baseline
Week 203The citation standard your briefs are held to
→Claim mapping, citation binding and the automation boundary
Week 104Access to relevant APIs, feeds or exports
→Public, licensed and internal corpus assessment, then integration setup
Week 205Briefs you would not want checked
→Retrieval cases and failure-mode testing
Week 406What no research brief may guarantee
→Retrievability scoring, review routing, guardrails and release controls
Week 307The researcher who will accept each claim
→Researcher release, then pilot and production validation
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Bands follow the effort a phase actually costs, and that is why one of them comes out carrying two.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Brief workflow discovery, claim mapping and the automation boundary
W2Corpus integration and the retrievability baseline
W3Claim assembly, retrieval logic and release controls
W4Evaluation suite, retrieval cases and failure-mode testing
W5Library integration, pilot claims and targeted corrections
W6One research quarter run under the research lead, then Agent Care handover
- Reading the band
- Each bar covers only the weeks its own work is named for, and one of them has to hold a second phase.
- At the end of W6
- Once the sourcing record validates, Agent Care assumes the agent.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Strategy AI agent
Build a deep research agent around the footnote nobody has ever opened.
Show us one claim in your last brief and the citation under it. Send that brief over and we will open every reference in it, mark which ones are primary and tell you which resolve to nothing. We set out what we searched; we do not claim to have found everything.
- Typical delivery
- 4–6 weeks
- Deployment
- Your last brief
- Researcher-held
- Retrieval-first
- After launch
- Agent Care available
Nestack Agents · deep research
AGT-CS-07 · Agent Care available after launch