This case uses a noisy, multi-tool, long-chain marketing research scenario to show how a harness controls execution and how latency, coverage, cost, evidence, and trace metrics prove delivery.
The marketing research scenario has noisy data, multi-step tool calls, evidence quality checks, and reusable public-demo requirements. The harness becomes the control layer for models, tools, state, evaluation, and replay.
The original issue was scattered sources, vague model summaries, weak traceability for client reviewers, and no process chain that could be explained in a live demo.
The delivery produced a live demo, eight harness nodes, evidence lineage, quality checks, and a replay document that clients can inspect and the team can rerun and extend.
Models change and tools get replaced; the harness determines whether the system can be delivered. It defines state flow, tool timing, evidence gates, output evaluation, and failure fallback.
The model handles local reasoning while the harness owns stage state, input-output contracts, tool order, and failure handling across long-running agent chains.
Marketing signals require multi-step handling. The harness turns external sources, evidence gates, strategy mapping, and delivery packaging into replaceable modules.
The same brief can be rerun and compared on source coverage, evidence density, strategy consistency, and output quality, creating room for continuous improvement.
Privacy, public safety, missing data, weak evidence, and unstable model behavior need gates. A harness turns those boundaries into enforceable system behavior.
Creator clips, hooks, recurring objections
LiveThreads, comparison language, trust markers
FilteredTexture, price, repeat-buy, failure modes
ScoredThis project measures more than “8 agent nodes.” The professional metrics cover end-to-end latency, source coverage, evidence coverage, duplicate-signal compression, planner redundancy, tool call success, human intervention, recovery, regression failure, cost per strategy pack, and trace coverage.
Full chain time from accepted brief to strategy pack, evidence summary, and demo talk track, covering the model call and surrounding workflow.
Validation: 20 fixed briefs, 5 public-source classes, 8-node harness, cold and warm cache split.Each run covers social, community, creator, review, and public-web signals so strategy draws from multiple source classes.
Validation: count valid source classes; duplicate and untraceable sources are excluded.Strategic claims, opportunities, and channel recommendations need public-signal or human-input support alongside model prose.
Validation: every key claim needs at least one reviewable evidence item or it counts as unsupported.Noisy research contains repeated comments and copied phrasing; the system compresses similar evidence before model reasoning.
Validation: compare evidence item count before and after dedupe while preserving source lineage.Includes brief normalization, collection planning, evidence filtering, pattern mapping, strategy generation, and handoff packaging.
Validation: tokens / successful strategy pack, strong model call ratio, and cache hit rate.All 8 agent nodes, tool inputs and outputs, evidence gates, and failure reasons must enter trace for replay and optimization.
Validation: each run can be replayed by node, tool, evidence gate, and final artifact.Research agents can repeatedly search, resummarize, and wander. Planner metrics constrain unnecessary execution steps.
Validation: compare actual steps with a human-labeled shortest effective path; repeated search, repeated filtering, and no-new-evidence steps count as redundant.Search, collection, filtering, evidence storage, and delivery packaging are all tool-chain work; failures affect the strategy pack.
Validation: count HTTP, parsing, schema, permission, and timeout errors per tool call; retry success is marked separately.The agent must call tools and construct correct queries, source scopes, filters, and output schemas.
Validation: trace sampling checks query, source class, time range, dedupe key, and output schema against task intent.High-risk briefs, unclear privacy boundaries, and weak evidence should enter human review while normal tasks keep automated flow.
Validation: human review / successful strategy pack, attributed to privacy, low evidence, strategy conflict, or client confirmation.When sources fail, tools time out, evidence is weak, or model output is invalid, the system should degrade, retry, or switch paths.
Validation: among recoverable failures, share that ends in a valid strategy pack or explicit human handoff; silent failures count as failures.Prompt, tool, source planner, or evidence-gate changes are checked against old briefs to protect strategy quality.
Validation: every release reruns a fixed brief set and checks evidence coverage, strategy consistency, human gates, and cost drift.The system turns one-off strategy prompting into an observable multi-node workflow with state, artifacts, and boundaries.
Key strategy claims stay attached to public-source signals, making output easier for clients to review, challenge, and calibrate.
Collection, evaluation, mapping, strategy, and packaging become reusable architecture rather than a one-off interface.
Turns product, goals, budget, and boundaries into stable project state.
Decides what can safely enter public demo and downstream reasoning.
Separates channel preference, sample density, and query strategy.
Connects social, community, creator, review, and web signals.
Filters duplicate, low-relevance, low-density, and unusable evidence.
Maps scattered comments into scenarios, motives, and audience structure.
Builds opportunity matrices, channel priorities, and strategic direction.
Packages strategy summary, evidence notes, and next-action checklist.
For website visitors: explain the problem, transformation path, and reusable architecture without exposing private details.
The standalone subdomain carries the complete product-like experience so clients can inspect workflow, nodes, evidence, and outputs.
Turns agent nodes, state, evidence contracts, evaluation scope, and safety boundaries into rerunnable governance assets.
The public page shows only client-verifiable content: how the brief enters the system, how agents execute, where evidence comes from, how weak outputs are gated, and how final artifacts are accepted.
Metrics constrain runtime quality. Every prompt, tool, source-planning, or evidence-gate change should rerun a fixed task set and inspect planner redundancy, tool calls, human intervention, recovery, and cost drift.
The case page explains metric scope and governance boundaries; the demo shows the full execution chain.
