{"slug":"research-harvey-case-study","name":"Case Study: The Harvey Harness","rules":{"start":"open","maxDepth":12,"allowRewind":true,"allowReplay":true},"nodes":{"open":{"short":"the note","pose":{"base":"centre"},"beat":{"layout":"center","marker":"case study · interactive","title":["Case Study:","The Harvey Harness"],"glass":1,"lede":"Reverse-engineered from Harvey's public technical disclosures: the advantage is the harness, not exclusive access to a smarter model. Harvey reports its auto routing cuts inference cost three to five times against a frontier-only baseline. A reconstruction, not a measurement.","caption":"12 beats · 15 min to read in full"},"prompt":"Where do you want to start?","choices":[{"label":"From the top","to":"premise"},{"label":"what is documented","to":"what-is-documented-what-is-infer"},{"label":"Straight to the end","to":"close"}]},"premise":{"short":"the premise","pose":{"base":"flooded"},"beat":{"layout":"left","marker":"01 · case study","title":["A claim runs","through this series"],"lede":"A claim runs through this series. The environment an agent works inside decides more than the model that executes the work. Our own experiments tested that on small tasks with our own runs.","caption":"Condensed from the note."},"prompt":"Where next?","choices":[{"label":"why harvey's edge","to":"why-harvey-s-edge-isn-t-a-secret"},{"label":"Wrap it up","to":"close"}]},"why-harvey-s-edge-isn-t-a-secret":{"short":"why harvey's edge","pose":{"base":"drained"},"beat":{"layout":"right","marker":"02 · why harvey's edge","title":["Why Harvey's edge isn't","a secret better model"],"lede":"The easy story about a company like Harvey is that it must have a private pipeline to a smarter model, a better Claude or GPT that the rest of us cannot buy. That story is wrong, and the architecture says so plainly.","caption":"Condensed from the note."},"prompt":"Keep going?","choices":[{"label":"not a wrapper","to":"not-a-wrapper-an-operating-syste"},{"label":"Wrap it up","to":"close"}]},"not-a-wrapper-an-operating-syste":{"short":"not a wrapper","pose":{"base":"centre"},"beat":{"layout":"left","marker":"03 · not a wrapper","title":["Not a wrapper,","an operating system"],"lede":"The word \"wrapper\" gets used for anything sitting between a user and a model. It is the wrong word here. Harvey is closer to a specialized, multi-agent operating system, and the foundation model is one commoditized device inside it.","caption":"Condensed from the note."},"prompt":"What next?","choices":[{"label":"the architecture","to":"the-architecture-in-layers"},{"label":"Wrap it up","to":"close"}]},"the-architecture-in-layers":{"short":"the architecture","pose":{"base":"flooded"},"beat":{"layout":"right","marker":"04 · the architecture","title":["The architecture,","in layers"],"lede":"Reconstructed from the disclosures, Harvey stacks into six layers. The bottom one is the commodity. The five above it are the product.","caption":"Condensed from the note."},"prompt":"Where next?","choices":[{"label":"the mechanisms","to":"the-mechanisms-and-the-weakness"},{"label":"Wrap it up","to":"close"}]},"the-mechanisms-and-the-weakness":{"short":"the mechanisms","pose":{"base":"recede"},"beat":{"layout":"left","marker":"05 · the mechanisms","title":["The mechanisms,","and the weakness each removes"],"lede":"Each mechanism exists to compensate for a specific weakness of the raw model. Read the table as a list of raw-model failure modes and the deterministic machinery Harvey built to remove each one.","panelRows":[{"title":"Dynamic routing (Auto)","note":"What it does Classifies each sub-task and routes it to the cheapest model that clears the Elo bar · The raw-model weakness it removes Frontier-only inference is uneconomic at hundreds of tool calls per run; vendor lock-in"},{"title":"Agentic Search (ReAct)","note":"What it does Runs 3 to 10 retrieval loops with a completeness check before it answers · The raw-model weakness it removes Single-pass RAG fails silently on a weak match and invents citations"},{"title":"In-memory OOXML editing","note":"What it does Model edits a minimal text view; deterministic code emits the diff and tracked changes · The raw-model weakness it removes LLMs corrupt document structure when asked to emit DOCX or XML directly"},{"title":"Spectre: durable runs, ephemeral workers","note":"What it does State is checkpointed to a durable control plane; workers spin up, write state, and die · The raw-model weakness it removes Context-window collapse on long tasks; ZDR violations from persistent third-party runtimes"}],"caption":"Condensed from the note, which carries 7 rows."},"prompt":"Keep going?","choices":[{"label":"what is documented","to":"what-is-documented-what-is-infer"},{"label":"Wrap it up","to":"close"}]},"what-is-documented-what-is-infer":{"short":"what is documented","pose":{"base":"drained"},"beat":{"layout":"right","marker":"06 · what is documented","title":["What is documented,","what is inferred"],"lede":"Honesty about evidence is the same discipline our experiments hold to, so the labels matter. Some of this is stated outright in Harvey's write-ups. Some is a strong inference from what a system with these properties must contain.","panelRows":[{"title":"A model abstraction layer normalizes tool-calls, stop conditions, and streaming across providers","note":"Strength Documented"},{"title":"Zero data retention rules out persistent managed runtimes, which is why Spectre uses ephemeral workers","note":"Strength Documented"},{"title":"Agentic Search runs iterative ReAct loops with a completeness check, not single-pass RAG","note":"Strength Documented"},{"title":"Vector storage is isolated per tenant, and post-retrieval filtering is rejected","note":"Strength Documented"}],"caption":"Condensed from the note, which carries 10 rows."},"prompt":"What next?","choices":[{"label":"why this beats raw","to":"why-this-beats-raw-model-access"},{"label":"Wrap it up","to":"close"}]},"why-this-beats-raw-model-access":{"short":"why this beats raw","pose":{"base":"flooded"},"beat":{"layout":"left","marker":"07 · why this beats raw","title":["Why this beats","raw model access"],"lede":"Set Harvey beside the three things people reach for instead, and the gaps line up with the mechanisms above.","caption":"Condensed from the note."},"prompt":"Where next?","choices":[{"label":"what harvey","to":"what-harvey-confirms-about-the-e"},{"label":"Wrap it up","to":"close"}]},"what-harvey-confirms-about-the-e":{"short":"what harvey","pose":{"base":"drained"},"beat":{"layout":"right","marker":"08 · what harvey","title":["What Harvey confirms about","the environment thesis"],"lede":"The reason this case study sits in the experiments section is that Harvey is the enterprise-scale version of what our own runs kept finding.","caption":"Condensed from the note."},"prompt":"Keep going?","choices":[{"label":"scope and limits","to":"scope-and-limits"},{"label":"Wrap it up","to":"close"}]},"scope-and-limits":{"short":"scope and limits","pose":{"base":"centre"},"beat":{"layout":"left","marker":"09 · scope and limits","title":["Scope","and limits"],"lede":"The caveats shape the claim, so they belong in the body.","rows":[{"title":"This is a reconstruction, not a measurement. ","note":"Every number here comes from Harvey's own disclosures, not from our rig. We did not run Harvey, and we cannot independently verify its benchmark figures."},{"title":"The inferences are labeled for a reason. ","note":"The reconciliation engine, the DAG compiler, and the synthetic-data pipeline are strong inferences, not disclosures."},{"title":"Legal is a specific domain. ","note":"Some of what makes the harness so heavy (the ethical walls, the zero-retention mandate, the disbarment-level cost of a permission slip) is specific to law."},{"title":"We are describing a moving target. ","note":"Product architecture at this stage changes on a quarterly cadence. This is Harvey as its public record described it in mid-2026, not a permanent account."}],"caption":"Condensed from the note."},"prompt":"What next?","choices":[{"label":"what transfers","to":"what-transfers-to-other-domains"},{"label":"Wrap it up","to":"close"}]},"what-transfers-to-other-domains":{"short":"what transfers","pose":{"base":"flooded"},"beat":{"layout":"right","marker":"10 · what transfers","title":["What transfers","to other domains"],"lede":"Strip the legal specifics and a blueprint is left for building an agent harness that beats native model chat in any domain where correctness and compliance matter more than conversational speed.","rows":[{"title":"Demote the model to a scheduled resource. ","note":"Normalize provider differences behind an abstraction layer, then route on cost and quality against real benchmark data."},{"title":"Make the fragile things deterministic. ","note":"Do not ask a model to emit a complex file format, reconcile parallel edits, or invent a citation."},{"title":"Separate transient state from durable state. ","note":"Ephemeral workers plus a durable control plane let a multi-hour task survive without a persistent process and without leaving privileged data at rest."},{"title":"Enforce permissions at the data layer. ","note":"Isolate storage physically per tenant and per project. Retrieving everything and filtering in application code is one logic bug away from a leak."}],"caption":"Condensed from the note, which carries 5 points."},"prompt":"Where next?","choices":[{"label":"Close the note","to":"close"}]},"close":{"short":"the end","pose":{"base":"finale"},"beat":{"layout":"center","title":["That was one path.","The note has the rest."],"glass":1,"lede":"This walk is one route through Case Study: The Harvey Harness. The full note carries every chapter, the tables, the code, and the source it was adapted from.","links":[{"href":"/research/harvey-case-study","label":"Read the full note"},{"href":"/research","label":"All research","tone":"ghost"}]}}}}