Agent-graph research workflow
Source workflow
Draw the real dependencies
Use one agent when every step needs the full result of the previous step. Use a graph when independent research angles can run together and later work genuinely reads their outputs.
Draw every job and edge. Keep an edge only when output from the first job passes into the second. Remove false waits such as summarizing a file before an unrelated calendar check.
Separate research from checking
Give each independent angle to a delegated Hermes sub-agent. Add a separate skeptic sub-agent that did not produce the original findings. Give the skeptic a narrow checking question; a vague request to "review the report" is not enough.
Have the skeptic read the findings and drop anything it cannot support before the merge. This separation matters more than adding more agents or choosing a secret model.
Merge and retain the last human yes
Merge the surviving findings into one report. Put human approval where a mistake becomes expensive to undo, usually before work leaves the business or moves money. Do not insert approval at every intermediate step.
Account for graph cost
Use cheaper models for broad collection and reserve the expensive model for judgment or the final merge. Preserve the source's caveat: one developer measured fixed overhead at 73% of each API call on one older Hermes setup; it is not a universal benchmark.
Build the first graph by interviewing the operator
Before building, ask:
- what question is being researched;
- which angles are worth splitting;
- what counts as a credible source;
- where the result should land.
Create one parallel sub-agent per angle, a separate skeptic that tries to disprove every finding and drops unsupported material, and a merge that ranks survivors by confidence with a source and date. Save the graph as a reusable skill, but do not schedule it yet.
Run it by hand until the operator can inspect what the skeptic drops and what the merge keeps. Once reliable, schedule the source example for Monday at 9am and send the finished report to Telegram. Send nothing else without asking.
Hermes is the source's durable environment: cron starts the graph, Telegram or Discord receives the report, and its delivery ledger preserves completed responses across a gateway crash.
XSkills operational adaptations
Preflight Hermes delegation, shared context, skill saving, cron, the delivery channel, and durable delivery. If delegation is unavailable, return the graph specification and prompts instead of claiming parallel execution. If scheduling or delivery is unavailable, complete the manual proof run and return the exact schedule/delivery configuration still needed.
XSkills safety adaptations
Use only authorized source access. Require the operator's approval before any external delivery, schedule activation, spending increase, or irreversible action. A research report may be prepared before that gate; it must not be sent through an unapproved channel.
Return
Return the graph or graph specification, angle findings with sources and dates, skeptic removals, ranked merged report, manual proof result, and any unactivated schedule or delivery step.
Source attribution
Independently compiled by XSkills from Tom Crawshaw's X Article about graph engineering with Hermes. Crawshaw did not author or review this skill. Capability fallbacks and explicit access, delivery, schedule, spending, and irreversible-action boundaries are XSkills adaptations. Attribution does not imply endorsement.