Automation candidates
While the Phase 8 catalog was being rebuilt, a parallel agent mined the local session corpus for work worth automating. It returned twenty-five ranked candidates.
The source report is coral-schema/automation-candidates.md in the audit
repository. This page reproduces its findings and explains the one structural
decision that makes it useful.
What was mined
| Source | Count |
|---|---|
| Session-state directories scanned | 2,923 |
Titled session plans (plan.md) | 262 |
| Machine-level skills | 100+ |
| User-level skills | 45 |
Fleet agents (dayour-*) | 80 |
| Ranked candidates returned | 25 |
The corpus is one workstation's accumulated agent sessions, not an organisation-wide survey. Candidates reflect what this operator repeatedly did by hand, which is exactly the signal that matters for deciding what to automate next, and exactly the wrong signal for claiming general demand.
The dedup baseline is the important part
Before ranking anything, the harvest established what was already being built, and flagged every candidate touching that set:
| Existing work | Status |
|---|---|
| GitHub key management | in flight |
owcp-incident-loadtest - timer-driven Dataverse incident generator | BUILT |
coral-agent-builder - one-shot create / name / instruct / greeting / publish / preview | in flight |
mcp-edge-unlock - clear a stale mcp-msedge lock to recover Playwright | in flight |
dataverse-rest-kit - az-token plus Web API metadata / create / update / file-column helpers | in flight |
flow-clone-runcontext - guided Save-As plus trigger Scope (User/Org) and Run-as | in flight |
darbot-file-validate - validate local file:// HTML via darbot-browser MCP | in flight |
pa-flow-pdf-to-record - persist a generated document back to its source row and flag it | in flight |
A ranked list of automation ideas is close to worthless without this step. Any sufficiently large session corpus will surface the same high-frequency tasks repeatedly, and those are precisely the tasks someone has already started automating - which is why they recur. Ranking by frequency alone reliably promotes solved problems to the top.
Four of the top ten overlap existing work. They stay on the list because overlap is not duplication - a partially-built tool may still need the generalisation a candidate describes - but they are marked, and they are not counted as net-new value.
Top ten
| # | Candidate | Rationale | Overlap |
|---|---|---|---|
| 1 | Bulk Agent Security Hardening | Apply web-channel and security settings across an agent fleet in one pass. | net-new |
| 2 | Eval-at-Publish ALM Gate | Run evaluations at publish time and block on regression. | net-new |
| 3 | One-Shot Coral Agent Provisioner | Parameterized create, configure, publish, and preview in a single pass. | coral-agent-builder |
| 4 | Dataverse REST Kit / Row-from-Template | Token-auth metadata plus create and update, including file columns. | dataverse-rest-kit |
| 5 | PDF/Doc to Record Persister | Write a generated document back to its source row and flag it. | pa-flow-pdf-to-record |
| 6 | Workshop Environment Pool Provisioner | Throttled parallel environment creation for labs. | net-new |
| 7 | HTML Wiki / Showcase Builder | Self-contained branded HTML wikis from a repository or session corpus. | net-new |
| 8 | GitHub Repo Bootstrap | Standardize CI/CD, issues, and Actions for new repositories. | net-new |
| 9 | Engagement / RoB Report Generator | Turn a CRM URL into an enriched HTML engagement report. | net-new |
| 10 | Screenshot/Browser Capture to Doc Pipeline | Privacy-safe captures feeding runbooks and wikis. | partial |
The strongest net-new additions, per the report's own summary: #1 Bulk Agent Security Hardening, #2 Eval-at-Publish, #6 Workshop Pool Provisioner, #7 Wiki Builder, #8 Repo Bootstrap, #9 RoB Report.
Two sketches worth reading
The report expressed its top two candidates as Coral workflow shapes, which is what makes them directly actionable against the node inventory in Phase 7:
Bulk Agent Security Hardening
Trigger -> Connector(Dataverse: list agents) -> Loop -> Switch(channel)
-> Connector(set security) -> Human review -> Note(audit) -> End
Eval-at-Publish ALM Gate
Trigger(publish) -> Function(run eval set) -> Classify(pass/fail) -> If/Else
-> [fail] Human review -> Note(scorecard) -> End
Both use node types this audit confirmed exist - Loop, Switch, If/Else, Function, Human review, Note. Neither has been built.
The second one is pointed directly at a failure this audit documented: Raybot's evaluation run scored 0% against questions that ground correctly in Preview. An eval gate built on that runtime today would block every publish. The candidate is sound; its dependency is not ready.
Candidates 11 through 25, by domain
Copilot Studio agent operations
| # | Candidate | Overlap |
|---|---|---|
| 11 | Agent Clone / User-vs-Org Parallel Runner | flow-clone-runcontext |
| 12 | Eval Scenario Library Expander | net-new |
| 13 | Adaptive Card / Flashcard Generator | net-new |
Dataverse / Power Platform
| # | Candidate |
|---|---|
| 14 | Choice / Option-Set and Metadata Discovery Helper |
| 15 | Power Automate Flow View / Pipeline Generator |
| 16 | PAC / Auth Session Manager |
M365 / Graph
| # | Candidate |
|---|---|
| 17 | Expense Report Drafter |
| 18 | Tenant Intelligence / Entra Enrichment |
| 19 | Lokka-Powered Graph Sync |
GitHub / DevOps
| # | Candidate |
|---|---|
| 20 | Repo / Codebase Audit and Analysis Report |
| 21 | Docusaurus / Docs-Site Publisher |
Documentation and deck pipelines
| # | Candidate |
|---|---|
| 22 | PPTX / Deck Builder (brand-aware) |
| 23 | DOCX / Architecture Doc Generator |
Knowledge graph and session intelligence
| # | Candidate |
|---|---|
| 24 | Session-History Harvester |
| 25 | Voice-of-Customer Aggregator |
Azure and infrastructure were assessed as lower priority for Coral specifically: cost optimization reporting, observability and KQL workbooks, and passwordless Postgres setup.
Candidate 24 is the harvester that produced this list. It nominated itself, and the report says so explicitly rather than quietly omitting it - a self-nominating tool is a legitimate finding about the corpus, and hiding it would have understated how much of this work is meta.
Candidate 21 - a docs-site publisher - describes this wiki. The wiki was built by hand before the harvest ran.
What this list does not establish
It is a ranking of opportunity, not a plan. Nothing here carries an estimate, an owner, a dependency graph, or evidence that the underlying platform can support it - as the Eval-at-Publish dependency above demonstrates.
The corpus is also single-operator, so "high value" here means "recurred often in one person's work." Treat the ranking as a prompt for a prioritisation conversation rather than its conclusion.
Related
- Phase 8 - Showcase catalog - built in parallel with this harvest
- Session forensics - the Copilot CLI session record for this repository
- Scripts - the automation actually present in the repository