Herendil Book a call All directions

trace engagementservice herendilrunning

We build AI agents that know how your company works.

Herendil is a small applied-AI consultancy. We find the work an agent can take on, connect it to what your company already knows, and hand it over with tests your team can run without us.

Book a callRead the trace
trace summary6 spans
  1. understandwk 1
  2. mapwk 1–3
  3. buildwk 3–7
  4. measurewk 4–9
  5. hand overwk 9–10
  6. awaiting inputnow
A typical shape. We size each engagement to the job.

understand.companyokwk 1

Understand the work

We spend the first week with the people who do the work. By the end we have a short list of jobs where an agent would save real time, and a clear view of what could go wrong in each.

  1. thinkStart from the work people repeat every week. Pick models later.
  2. callinterview(team="support")
  3. callinterview(team="operations")
  4. readrunbooks/, ticket history, shared drives
  5. thinkMost of this job is looking things up. The judgment calls are rarer and easy to spot.
  6. okshortlist: jobs ranked by time saved and risk

map.knowledgeokwk 1–3

Map what the company knows

An agent can only use what it can find. We map where knowledge lives, who is allowed to see it, and which questions each source answers. This is the start of a company brain: what your company knows, decides and does, available to people and to agents.

  1. callindex(source="wiki")
  2. callindex(source="tickets", since="2y")
  3. callindex(source="chat", scope="public channels")
  4. callresolve(permissions) → agents see what the asker may see
  5. warnThe refund policy lives in three places, and two of them disagree.
  6. okmap: sources × questions, each with an owner

build.agentokwk 3–7

Build the agent

We build the smallest agent that does the job, inside tools your team already uses. Each step gets the model that suits it, and inference runs where your cost and data rules allow, whether that is a hosted API, your cloud or your own hardware.

retrieve410ms plan260ms crm.lookup620ms check.policy340ms draft.reply580ms
  1. thinkThe agent drafts. A person approves before anything reaches a customer.
  2. callretrieve(question, map="company")
  3. callcrm.lookup(customer_id)
  4. callcheck.policy(action="refund")
  5. okdraft ready for review, sources attached

measure.outcomesokwk 4–9

Measure it on real cases

Before anything ships, we turn past cases from your own work into tests. Every change runs against them, so you can see whether the agent got better or worse, and on which cases.

  1. calleval.run(set="past cases")
  2. failrefunds over the limit escalate too late
  3. thinkThe failures share one cause: the conflicting policy page. Fix the source, then run again.
  4. calleval.run(set="past cases")
  5. okno regressions, report sent to the team

handover.teamokwk 9–10

Hand it over

Your team owns the result. The code, prompts, tests and runbook live in your repositories. We work alongside the people who will run it until they no longer need us.

  1. callgit.push(repo="yours")
  2. calldocs.write("runbook.md")
  3. callpair(team="platform", until="confident")
  4. okowner: your team

emit.notesokongoing

Notes

We write down what we learn along the way.

await.inputwaitingnow

Book a call

Tell us which work you would hand to an agent. In 30 minutes we will tell you whether an agent fits and what the first week would look like.

herendil › awaiting input from you

Book a call

Online scheduling opens soon.