Skip to content
AADM
MCP

▸ 29 vignettes · agent sessions with the hosted AADM MCP

Patterns from the field.

Portfolio band

~899–1,375engineer-hours

Redirected from wrong fixes, duplicate UI, false CI, and meeting loops across these 20+1 patterns — not raw codegen speed.

These are real session patterns from teams running the AADM MCP. The app names are fictional. The bugs, refactors, and decisions are not. Hours are illustrative — the typical-wrong-path vs the attested-path, for a senior engineer + agent pair.

Every vignette below is the same move: a senior engineer's judgment applied before the agent starts editing rather than after. Agentic engineering is the general form of that move.

Read the vignettes Get access The honest boundary

Fictional apps. Real patterns.

A

App names are fictional

StudioPublish, LedgerFlow, FleetDispatch — none of these are real customers. The bugs, refactors, and product calls behind each story are real; the brand has been anonymized.

B

Hours are illustrative

Ranges compare a typical wrong-path (meetings, false fixes, full CI, parallel UI) against the attested AADM path. A senior engineer + agent pair, not a guaranteed SLA.

C

AADM frames; your repo proves

AADM MCP frames problems, names seams, and enforces evidence obligations. Your repository's tests and CI still decide pass/fail. Nothing here is a compliance verdict.

The catalog. In one place.

Each story walks through the same four beats: the situation your team is in, the decision you face, what AADM had the agent do, and what changed. App names are fictional; the patterns are real. Skim for the one that looks like your week.

  • 01

    Media

    StudioPublish

    12–18 hrs saved

    Two exits, not one Back button

    The situation
    Your producers open an export page from three places: a campaign, the review queue, or the library. One Back button sent them to the wrong place.
    The decision
    Patch the header button, or decide where each person actually needs to go.
    What AADM did
    The agent walked the screen from the producers' own words, mapped every way in and out, and stress-tested the plan before any code.
    The impact
    Two clear exits, back to where you were and back to the hub, with the tests named before anything shipped.

    Three AADM tools that did the work

    • Walk the screen as your user ux_surface_pass
    • Map every move and the call behind it user_flow
    • Stress-test the plan pressure_test
  • 02

    Edtech

    Vertex Learning

    20–30 hrs saved

    Find the integration gaps

    The situation
    Product promised one publish flow across slides, voice, video and the LMS. Engineering had half-built background jobs and fourteen scattered tickets.
    The decision
    Keep holding cross-team workshops, or get one honest list of what works and what doesn't.
    What AADM did
    The agent traced the ten-step author journey and marked each step as proven, unproven or not built, including what operators could monitor.
    The impact
    One signed-off requirements brief replaced the workshops.

    Three AADM tools that did the work

    • Walk the screen as your user ux_surface_pass
    • List what operators need to see signal_stack_outline
    • Mark each step proven or not journey_attest
  • 03

    Retail

    Northstar Retail

    30–45 hrs saved

    Split a 4,000-line screen safely

    The situation
    One merchant admin screen, about 4,000 lines, shared state and cache refreshes across three panels. A careless split would have broken who is allowed to purge data.
    The decision
    Rewrite it in one go, or split it in an order that keeps permissions intact.
    What AADM did
    The agent sequenced the split, pulled in the tenant permission rules that applied, and challenged the purge-versus-archive design before code moved.
    The impact
    Small, reviewable slices shipped one at a time. Nothing was deleted and hoped for.

    Three AADM tools that did the work

    • Sequence the work plan_delivery
    • Pull in the controls that apply fetch_auth
    • Challenge the design architecture_lens
  • 04

    Analytics

    Pulse Analytics

    16–24 hrs saved

    Green tests, wrong numbers

    The situation
    Every test passed, yet finance disputed the usage percentages. The client was inventing fields and the test mocks used the wrong metric name.
    The decision
    Trust the green dashboard, or check that the tests measure the real thing.
    What AADM did
    The agent separated “is the data right” from “is it measured right,” then forced an explicit pass or fail on each data contract.
    The impact
    The false-green tests were exposed and fixed on their own track, apart from the real data work.

    Three AADM tools that did the work

    • Pin down the facts first prism_frame
    • Force pass or fail at each boundary sentinel_closure
    • Mark each step proven or not journey_attest
  • 05

    Fintech

    LedgerFlow

    24–40 hrs saved

    Billing cycle math

    The situation
    Month-end credits overshot because the query counted the last day of the cycle when the contract excluded it. One webhook path skipped balance checks entirely.
    The decision
    Patch the numbers on the dashboard, or fix the ledger at the source.
    What AADM did
    The agent stayed evidence-first: reproduce the overshoot, find the exact query, and map every place money gets reconciled.
    The impact
    The ledger was fixed at the source. No dashboard band-aids, no “upgrade the database” detour.

    Three AADM tools that did the work

    • Investigate before fixing debug_defect
    • Pin down the facts first prism_frame
    • Mark each step proven or not journey_attest
  • 06

    Insurance

    Harbor Insurance

    8–14 hrs saved

    The missing CORS header

    The situation
    Members saw empty account settings in production while integration tests passed. A custom request header was missing from the server's allow-list.
    The decision
    Add timeouts and retries, or prove exactly where the request dies.
    What AADM did
    The agent followed the evidence in order: the browser's network trace, the code that sets the header, the allow-list, then a contract test. It blocked the timeout guesses.
    The impact
    One fix, plus one test that keeps it fixed.

    Three AADM tools that did the work

    • Investigate before fixing debug_defect
    • Bundle the fix with its proof fix_pack
    • Pick the right workflow route_lane
  • 07

    Construction

    BridgeWorks

    10–16 hrs saved

    The missing row behind a 500

    The situation
    A new work package appeared, but its preview returned a server error. On-call wanted to double the query timeouts.
    The decision
    Treat it as an infrastructure problem, or find the logic gap.
    What AADM did
    Triage ruled out infrastructure. A targeted search found the step that skipped building the data the first time it was needed.
    The impact
    A scoped fix with no rollback and no timeout bump.

    Three AADM tools that did the work

    • Sort the problem triage_bug
    • Dig for the real cause digfind
    • Investigate before fixing debug_defect
  • 08

    Regtech

    Atlas Compliance

    12–20 hrs saved

    Noise, not an incident

    The situation
    Monitoring fired 258 critical alerts in one day. Leadership wanted to roll back the release.
    The decision
    Roll back, or first check what the alerts really mean.
    What AADM did
    The agent sorted the alerts: one self-healing backlog, duplicate worker messages, and zero stale data.
    The impact
    No rollback. The fix was alert hygiene, and customers were never affected.

    Three AADM tools that did the work

    • Sort the problem triage_bug
    • Name what kind of bug it is classify_bug
    • List what operators need to see signal_stack_outline
  • 09

    Logistics

    FleetDispatch

    25–35 hrs saved

    Three IDs, one column

    The situation
    Mid-migration, three kinds of ID existed at once. Export routes wrote the wrong one into database links, which failed silently.
    The decision
    Paper over it with fallbacks in code, or stop and decide.
    What AADM did
    The agent wrote down what was partly shipped, named the broken contract, and listed the four decisions product still owed.
    The impact
    The migration continued with a decision log instead of fallback code spreading through the app.

    Three AADM tools that did the work

    • Pin down the facts first prism_frame
    • Name the broken contract fetch_seam
    • Map the feature and its blockers map_feature
  • 10

    Healthcare

    Meridian Health

    14–22 hrs saved

    Tests that no longer exist

    The situation
    CI scripts pointed at test files that had been deleted, and the registry still said they ran. Auditors assumed coverage that wasn't there.
    The decision
    Keep shipping, or prove the registry matches the code.
    What AADM did
    The agent required proof that every referenced file exists before cutover, and kept the work scoped as cleanup rather than new features.
    The impact
    The registry matched the code again, and the audit claims were true.

    Three AADM tools that did the work

    • Bundle the fix with its proof fix_pack
    • Keep key rules true over time invariant_register
    • Pick the right workflow route_lane
  • 11

    HR tech

    ClearPath Talent

    18–28 hrs saved

    A demo that proved nothing

    The situation
    A static skills map proved nothing. Engineering proposed a browser-side chart with made-up weights, even though server APIs already held the real data.
    The decision
    Build the pretty demo, or wire it to real data.
    What AADM did
    The agent tied every control (expand, filter, cluster) to existing server data before any wireframe.
    The impact
    The feature used real numbers from day one. No throwaway demo.

    Three AADM tools that did the work

    • Walk the screen as your user ux_surface_pass
    • Map the feature and its blockers map_feature
    • Map every move and the call behind it user_flow
  • 12

    Media

    StudioPublish

    40–60 hrs saved

    A screen that said “ready” too soon

    The situation
    The publish dashboard crashed when stored data lagged the code reading it. Fixing the crash revealed a second problem: saves didn't refresh status, so the screen could say “ready” when it wasn't.
    The decision
    Ship the crash fix alone, or close both problems.
    What AADM did
    The agent investigated the crash, then traced every step that should refresh after a save.
    The impact
    Both problems were fixed in one release, and the screen stopped lying.

    Three AADM tools that did the work

    • Investigate before fixing debug_defect
    • Mark each step proven or not journey_attest
    • Bundle the fix with its proof fix_pack
  • 13

    Regtech

    Atlas Compliance

    32–48 hrs saved

    Stuck on “computing”

    The situation
    Audit summaries hung for over 20 minutes while the query itself took 5 milliseconds. Ten requests were rebuilding the same thing at once and fighting over locks.
    The decision
    Raise timeouts everywhere, or stop the duplicate work.
    What AADM did
    The agent named the cause, missing de-duplication, and listed the fixes that were off the table, including timeout increases.
    The impact
    One request builds and the rest wait. Summaries return in seconds.

    Three AADM tools that did the work

    • Investigate before fixing debug_defect
    • Name what kind of bug it is classify_bug
    • Apply the coding rules coding_principles
  • 14

    Edtech

    Vertex Learning

    35–50 hrs saved

    Saved, but empty

    The situation
    Lessons saved with a success message, but the plan panel stayed empty. The screen waited on an optional field while the data behind it failed on first load.
    The decision
    Pre-build everything (60 to 90 second hangs), or fake it in the browser.
    What AADM did
    The agent rejected both shortcuts and named the contract between saving and displaying.
    The impact
    Fixed at the boundary, before more features were stacked on top.

    Three AADM tools that did the work

    • Pick the right workflow route_lane
    • Name the broken contract fetch_seam
    • Map every move and the call behind it user_flow
  • 15

    Construction

    BridgeWorks

    28–40 hrs saved

    The phantom conflict

    The situation
    The first save of a brief claimed someone else had changed it. Timestamps lost precision in JavaScript, so versions never matched, and the recovery message told users to re-import their work.
    The decision
    Loosen the check, or change how versions are tracked.
    What AADM did
    The agent applied the coding rules and planned a move to whole-number revision counters with clear conflict messages.
    The impact
    No more phantom conflicts, and real ones explain themselves.

    Three AADM tools that did the work

    • Pick the right workflow route_lane
    • Apply the coding rules coding_principles
    • Sequence the work plan_delivery
  • 16

    Insurance

    Harbor Insurance

    24–36 hrs saved

    Show what's shipping

    The situation
    Underwriters saw pass and fail gates with no proof behind them. A new “what we're shipping” panel was about to merge without evidence.
    The decision
    Merge on trust, or attach the proof.
    What AADM did
    The agent checked every step of the release journey and signed off each boundary between teams.
    The impact
    Shipped with known gaps written down: nothing hidden, nothing blocking.

    Three AADM tools that did the work

    • Mark each step proven or not journey_attest
    • Force pass or fail at each boundary sentinel_closure
    • Attach proof to each claim repo_evidence_binder
  • 17

    HR tech

    ClearPath Talent

    45–70 hrs saved

    One assistant, not five

    The situation
    Recruiting, onboarding, learning and analytics each hard-coded their own assistant actions. A new console was about to add a fifth.
    The decision
    Add another one, or unify first.
    What AADM did
    The agent mapped the feature, planned one shared decision point, and questioned the plan until it held up.
    The impact
    One assistant logic, clear reasons when an action is dropped, and CI checks that keep it that way.

    Three AADM tools that did the work

    • Map the feature and its blockers map_feature
    • Sequence the work plan_delivery
    • Ask the hard questions grill_questions
  • 18

    Media

    StudioPublish

    30–45 hrs saved

    Don't build the menu

    The situation
    A spec still asked for a four-option dropdown, but a button strip already in production did the same thing on the same backend.
    The decision
    Build what the spec says, or check whether it adds anything.
    What AADM did
    The agent asked what would break if the menu never existed. Nothing would.
    The impact
    The spec was updated instead of building duplicate code.

    Three AADM tools that did the work

    • Ask the hard questions grill_questions
    • Challenge the design architecture_lens
    • Question the plan interrogate_plan
  • 19

    Edtech

    Vertex Learning

    16–24 hrs saved

    Settle the rules before automating

    The situation
    A background job was ready to generate course outlines beside the main system, giving two places that could make plans that disagree.
    The decision
    Wire it up now, or settle the rules first.
    What AADM did
    The agent ran a go/no-go review and held the job until one shared set of teaching rules existed.
    The impact
    One source of truth for course structure before anything ran unattended.

    Three AADM tools that did the work

    • Ask the hard questions grill_questions
    • Challenge the design architecture_lens
    • Sequence the work plan_delivery
  • 20

    Regtech

    Atlas Compliance

    50–80 hrs saved

    Stop running all of CI

    The situation
    Repairs touched more than 50 test files, and every guess triggered a 90-minute CI run.
    The decision
    Keep guessing, or narrow the target.
    What AADM did
    The agent marked which steps were proven and which weren't, then grouped the evidence by cluster.
    The impact
    Targeted test runs replaced a full CI run on every guess.

    Three AADM tools that did the work

    • Mark each step proven or not journey_attest
    • Bundle the fix with its proof fix_pack
    • Investigate before fixing debug_defect
  • 21

    Logistics

    FleetDispatch

    20–30 hrs saved

    Product decisions first

    The situation
    Engineering was ready to migrate, but product hadn't made four key decisions, and coders had started guessing.
    The decision
    Let coders decide, or put the decisions in front of product.
    What AADM did
    The agent broke the work into issues and wrote the open questions as a numbered menu.
    The impact
    Product answered in one session, and coders stopped guessing.

    Three AADM tools that did the work

    • Break work into issues slice_issues
    • Ask the hard questions grill_questions
    • Map the feature and its blockers map_feature
  • 22

    Marketplace

    Cartwheel

    60–90 hrs saved

    Untangle the vibe-coded app

    The situation
    Two founders prompted their marketplace into existence: 140 files, three copies of checkout, database tables anyone could read, and secret keys shipped to the browser. Every new prompt broke something old.
    The decision
    Rewrite from scratch, or rescue what's there.
    What AADM did
    The agent mapped what existed, checked which tables leaked between customers, and ordered the work: lock down data, merge duplicates, then features.
    The impact
    Twelve ordered fixes, no rewrite, and the store kept taking orders the whole time.

    Three AADM tools that did the work

    • Map the codebase repo_architecture_map
    • Check customer data stays separate tenant_isolation_check
    • Break work into issues slice_issues
  • 23

    Fintech

    LedgerFlow

    120–180 hrs saved

    Refactor the whole backend

    The situation
    Eight years of routes calling the database directly: business rules in controllers, validation in three places, and no part that could be changed safely.
    The decision
    A big-bang rewrite, or replace it route by route.
    What AADM did
    The agent planned a step-by-step move (input boundaries first, business logic next, data access last) and defined old-versus-new checks each route had to pass before traffic moved.
    The impact
    The backend moved in steps, with the money rules checked at every one.

    Three AADM tools that did the work

    • Sequence the work plan_delivery
    • Compare old and new before switching cutover_parity_check
    • Keep key rules true over time invariant_register
  • 24

    Proptech

    Tenantly

    50–80 hrs saved

    Prototype to production

    The situation
    The demo that won the pilot went live as-is: no error plan, no audit log, retries that charged rent twice, and nothing operators could debug at 2 a.m.
    The decision
    Keep patching, or prove it's ready.
    What AADM did
    The agent checked one area at a time against the standard's controls, built a single list of errors, and named what operators need to see.
    The impact
    “Ready” became a checklist with evidence attached.

    Three AADM tools that did the work

    • Prove readiness one step at a time assurance_run
    • Build one catalog of errors error_registry_check
    • List what operators need to see signal_stack_outline
  • 25

    SaaS

    Orbit Ops

    30–50 hrs saved

    The agent that could drop prod

    The situation
    An internal coding agent held a permanent database password and ran migrations on request, even during a code freeze.
    The decision
    Trust it to behave, or limit what it can do.
    What AADM did
    The agent mapped who the bot acts for, what it can reach, and how to shut it off.
    The impact
    Short-lived credentials per task, a freeze it can't override, and a kill switch tested before anyone trusted it again.

    Three AADM tools that did the work

    • Run a focused audit sentinel_program
    • Check what the agent may do agent_op_contract_check
    • Pause or stop an agent run run_state_control
  • 26

    Healthcare

    Meridian Health

    40–60 hrs saved

    FDE: tailor the core, don't rewrite it

    The situation
    A forward-deployed engineer had six weeks inside a hospital network to make your core referral backend fit their workflow. They had the hospital's source code, but rewriting it was never on the table.
    The decision
    Fork the product for one customer, or extend the core in place through configuration and plugins.
    What AADM did
    The agent mapped the hospital's code onto the UDALI layers, showed which seams already took configuration or plugins, and sequenced each customization against those seams with a named acceptor on the hospital side.
    The impact
    Hospital-specific rules shipped as config and plugins on the unchanged core. Their team owns them, and the reusable pieces went back into the product.

    Three AADM tools that did the work

    • Map the codebase repo_architecture_map
    • Sequence the work plan_delivery
    • Run a focused audit sentinel_program
  • 27

    Logistics

    Keystone Logistics

    30–45 hrs saved

    FDE: no source code, still structured

    The situation
    Your FDE landed at a freight customer whose core systems belong to a vendor. There was no source code to read: only API docs, runbooks, interview notes and a few exports.
    The decision
    Guess at how their system works, or build a shared, written picture of what is known.
    What AADM did
    The FDE connected AADM to a repo of what was known (docs, contracts, notes) and used the FDE tools to place each piece on the UDALI layers with the customer's architects, marking every gap as unknown instead of assuming.
    The impact
    Customer and FDE agreed on one map of their system, with the gaps listed, before any integration code was written. The plugins that followed targeted real seams.

    Three AADM tools that did the work

    • Sequence the work plan_delivery
    • Map one feature from what is known repo_feature_map
    • Map the feature and its blockers map_feature
  • 28

    Insurance

    Granite Mutual

    35–55 hrs saved

    FDE: tools the customer keeps

    The situation
    The customer wanted their own agents to quote and file claims against your core backend, safely, after the FDE left.
    The decision
    Hand over raw API keys and a wiki page, or ship governed tools the customer can run themselves.
    What AADM did
    The FDE used AADM to design an MCP server, a small set of APIs and agent skills on top of the core: checked each tool's schema across model providers, scoped what every agent may do, and wrote the proof each tool needed before release.
    The impact
    The customer runs their own MCP, APIs and skills on your core, with permissions and tests they understand and can change.

    Three AADM tools that did the work

    • Check tool schemas work across models llm_tool_schema_compat
    • Check what the agent may do agent_op_contract_check
    • Prove readiness one step at a time assurance_run
  • 29

    Retail

    Northstar Retail

    25–40 hrs saved

    Ten agents, one codebase

    The situation
    The team ran several coding agents at once and spent more time on merge conflicts than on features.
    The decision
    Run fewer agents, or split the work better.
    What AADM did
    The agent split work so no two agents touched the same files, checked each branch before merge, and recorded what each run got wrong.
    The impact
    Parallel work that merges cleanly, and agents that learn between runs.

    Three AADM tools that did the work

    • Split work across agents fanout_guide
    • Check each branch before merge chunk_branch_check
    • Record what each run got wrong lessons_register

Hour ranges are method · typical wrong path (meetings, wrong fixes, full CI, parallel UI) vs attested path (one governed agent session). A senior engineer + agent pair. Not a guaranteed SLA. Not a compliance claim. Your repository's tests and CI still decide pass/fail.

The portfolio band. All of them together.

If the patterns above all ran in your quarter — they won't all show up, but the shapes recur — the redirected effort adds up.

If these patterns landed in one quarter

~899–1,375

engineer-hours redirected

Redirected from wrong fixes, duplicate UI, false CI, and meeting loops. Not raw codegen throughput. Not a compliance claim. A summed estimate of the kinds of wasted time these patterns prevent.

By industry · 29 vignettes

  • Media 3
  • Edtech 3
  • Insurance 3
  • Regtech 3
  • Logistics 3
  • Retail 2
  • Fintech 2
  • Construction 2
  • Healthcare 2
  • HR tech 2
  • Analytics 1
  • Marketplace 1
  • Proptech 1
  • SaaS 1

How we talk about this. How we don't.

The framing rules we hold ourselves to when reporting these patterns. Useful if you're considering AADM and want to know what claims we'll make in front of your auditor.

We say

  • ✓ 12–18 hours avoided wrong Back button implementation and review cycle
  • ✓ debug_defect classified root cause in one session; forbade timeout padding
  • ✓ Round 2 journey_attest caught stale verdicts before ship
  • ✓ Vertical through_domain_and_data plus horizontal seams signed off

We don't say

  • ✗ AADM saved 90% of development
  • ✗ AADM fixed production
  • ✗ Zero bugs after AADM
  • ✗ Certified compliant

Honest boundary · AADM MCP frames problems, names seams, and enforces proof obligations. Your repository's tests and evidence still decide pass/fail. That's the line, and we hold it.

The best case study is yours.

Try the patterns on something real this week. Catch a classification bug, attest a release without theater, onboard a senior engineer in a week. When it works, tell us — we'll feature consenting teams here.

See how the MCP works Read the standard Why AADM, vs raw loops
Catalog · 29 vignettes
Hours band · ~899–1,375 engineer-hours redirected