Behavior Guidance Packs
The behaviors your agent has to get right.
You turn your expert method into testable behaviors and a quality loop, so every release of your coach, tutor, or advisor scores better than the last.
We refresh the packs as the evidence and the best thinking in your field move.
You're not starting from scratch
We bring the standard. You compound the advantage.
A behavior pack is a tested library we bring in on day one. It names the behaviors a coaching or advisory agent has to get right, with the failure modes and the pass and fail checks already worked out. We tune it to your product, and the tuning is where you build the advantage. We keep the reusable library, the taxonomy, personas, and scorers. You own the tuned datasets, the golden conversations, and your method layer.
Start from a tested behavior library
You skip the blank page. We bring the behaviors a coaching or advisory agent has to get right, drawn from the ways these products fail with real users, so yours clears the bar in week one instead of finding the gap after a user has already left.
Build evals only you have
We tune the behaviors to your method and your real conversations. You turn your hardest moments into a test set only you hold, adding cases and raising the bar every month you use it. A competitor cannot copy that private test set.
Catch the regression before it ships
Every prompt, model, and flow change runs against the pack first. The scorers are calibrated against your experts, so a green check means the agent does the right thing in the moments that decide trust. You see the regression in review, not in a user who quietly stops coming back.
Then give the agent a way to act. Specialized tools are the screens your agent opens when talking it through is not enough, like a plan it can save or a guided reset.
Explore specialized tools →What's inside a pack
Your eval platform tells you what happened. The pack defines what should have happened.
Your platform ships the machinery, the datasets, scorers, and dashboards, plus a blank space where the definition of good is supposed to go. For expert guidance, that definition is the hard part, and a generic template won't fill it. Each behavior here is one observable unit, with the failure it prevents, what passing and failing look like, the dataset behind it, and the scorer that runs it.
Behavior
reflect_before_advisingReflect before advising
Risk prevented. The agent sounds clinical, dismissive, or prematurely solution-oriented, and the user stops trusting it with anything hard.
Pass
It names the user's emotional state, reflects the real tension back, and asks one useful question before offering a single tactic.
Fail
It gives advice before it understands the felt conflict, stacking steps onto a person who needed to feel heard first.
Test scenarios
- Angry parent after a hard school call
- Ashamed manager who just missed a deadline
- Skeptical executive testing whether it's safe
- Overwhelmed employee venting before deciding
Tested against 24 cases across shame, anger, fear, and overwhelm.
Runs as
LLM-as-judge, reference-free · pass/fail · final response
A pack is runnable, not a doc
It compiles to the five objects an eval stack already ingests, and imports into Coval, Braintrust, LangSmith, or Langfuse.
Scenario datasets
Real critical moments, organized by failure mode, not generic prompts.
Personas
Reusable simulated users such as the ashamed high performer, the skeptical executive, the overwhelmed parent, and the boundary tester.
Scorers and judges
LLM and code scorers calibrated against expert judgment, each with labeled examples, a threshold, and the rule for when a human overrides the judge.
Trace and metadata contract
The fields your product must log, so a judge can see what happened, not just the transcript.
Release gates
What is good enough to ship, and what blocks a release even when task success looks fine.
The packs
Narrow on purpose, so they're hard to copy.
A platform can auto-generate a generic empathy or hallucination judge from a prompt in seconds. It can't generate the private context these encode, the real transcripts, your method, your risk tolerance, and the failure modes that only surface in emotionally loaded, expert-led work. That judgment is the product, and it's why this doesn't compress into a platform feature.
Critical Guidance Moments
The high-stakes moments where guidance products fail, when a user is activated, ashamed, or in real risk and there is no clean right answer. Get one wrong and the user stops bringing the hard things, then stops coming back.
Measures
- Reflect before advising
- Ask one useful next question
- Escalate without abandonment
- Avoid shame amplification
Method Fidelity
Whether the agent follows your method instead of drifting into generic LLM advice. Your method is the reason a user picks you over a free chatbot, and drift erases the one thing they can only get from you.
Measures
- Right intervention at the right stage
- Doesn't skip discovery
- Holds your distinctive voice
- Knows when the user isn't ready
Enterprise Trust & Reporting
Whether the product is safe to deploy inside an enterprise, where privacy and reporting are the real buying friction. This is the evidence that clears security review and unblocks the deal.
Measures
- Private content stays out of summaries
- Support, not surveillance
- In scope on HR, legal, and medical
- Clean escalation
How we bring it in
Audit what you have. Build the pack. Keep the loop.
Three ways to start, mapped to how we already work. Most teams begin with an audit of the agent they already have, then build the pack the gaps point to.
Agent Behavior Audit
You have an agent or prototype and something feels off, but you can't yet name it.
We run your transcripts, prompts, and risk profile against the behavior library, then hand back a gap map and a launch-readiness scorecard. You see where the agent stands today, with a scorecard you can put in front of an enterprise buyer or a security review.
What you leave with
- Behavior gap map
- Launch-readiness scorecard
- Prioritized risk list
- A recommended next step
Behavior Pack Buildout
You're ready to make the quality layer real before serious users depend on it.
In two to four weeks we deliver a runnable pack for one product area or launch risk, with datasets, personas, scorers, a trace contract, and release gates, plus one adapter for your stack.
What you leave with
- Platform-ready scenario datasets
- Persona library
- Calibrated scorers and judge prompts
- Trace and metadata contract
- Release gates and review guide
- One adapter for your stack (Langfuse, Coval, or Braintrust)
Quality Loop Subscription
You're in pilot or production and the agent has to keep getting better.
We turn production failures into new test cases, recalibrate scorers, tune thresholds, and run release review. When a new model ships, you run it against the pack and find out if it is better for your users before you switch. Every month your private test set grows and your bar sharpens, and that test set is the part a competitor cannot copy.
What you leave with
- Production failures converted to test cases
- Scorer recalibration
- Threshold tuning
- New models vetted against your bar
- Release-gate review
- A loop your team keeps
Quality you can take to a buyer.
You leave able to show a buyer why the agent is ready, what it won't do, and how you catch a regression before it ships, with the loop set up to keep improving. The same quality layer that clears enterprise review is what keeps users coming back.
In build. Design partners open.
Run the packs against live traces, not just at build time.
The packs are already runnable. hunter-guard makes them a package that plugs into Langfuse and the rest of your eval stack, so the same scorers run against live traces and tie back to the outcomes you track, not only at build time. We are scoping it now with a design partner. Ask if you want to shape it early.
- Packages the same calibrated scorers your pack already defines
- Plugs into Langfuse and the rest of your eval stack, no rebuild
- Ties eval results back to the outcomes you already track
Common questions
What teams ask about the packs.
A tested library of the behaviors a coaching or advisory agent has to get right, with the failure modes and the pass and fail checks already worked out. It compiles to scenario datasets, personas, scorers, a trace schema, and release gates, the five objects an eval stack already ingests.
A pack imports into Coval, Braintrust, LangSmith, or Langfuse. It compiles to the objects those platforms already run, so it plugs into the stack you have instead of replacing it.
A platform can auto-generate a generic empathy or hallucination judge from a prompt. It cannot generate the private context these encode, your real transcripts, your method, your risk tolerance, and the failure modes that only surface in expert-led work. That judgment is the part a competitor cannot copy.
You own the tuned datasets, the golden conversations, and your method layer. We keep the reusable library we bring in, the taxonomy, personas, and scorers. Every month you use it, you add cases to a test set only you hold.
Three ways, mapped to how we already work. An audit of the agent you already have, a buildout of the pack the gaps point to, or an ongoing quality loop. Most teams begin with the audit.
You can show a buyer why the agent is ready, what it will not do, and how you catch a regression before it ships. The same quality layer that clears procurement is what keeps users from churning.
That is hunter-guard, an SDK in build now. It packages the same scorers to plug into Langfuse and the rest of your eval stack, so the checks run against live traces in production, not only at build time. It is open to design partners while we scope it. Ask David for early access.