Routing local and hosted coding agents through isolated worktrees, escalation, independent review, and evidence-based acceptance.
Development and validation
Local coding agents
Hosted models
Git worktrees
Independent review
Automated tests
Implementation and acceptance are separate decisions. A failed or unavailable review is not approval.
My contribution
Developed coding-agent routing, worktree isolation, and review workflows.
Engineering decision
Have a separate reviewer check the coding agent’s work.
Evidence to explore
Tested local-only tasks, hosted-model escalation, and separate review steps.
Boundary
These tests cover specific workflows. Broader reliability and cost comparisons are still open.
01Route task
02Reuse worktree
03Implement
04Escalate if needed
05Verify independently
Why I built it
Herdr routes coding tasks between local and hosted agents. I built it to keep their changes in separate Git worktrees and have the results reviewed, instead of relying on an agent’s message that the task is done.
What I worked on
Route work between local and hosted coding agents while avoiding unnecessary hosted-model usage.
Reuse prepared Git worktrees to preserve isolated changes and keep primary working directories unchanged.
Separate implementation from independent review and preserve failure evidence.
Use automated test suites as acceptance evidence rather than treating the agent’s completion message as sufficient.
Representative validation workflows
Local-only work completed with zero hosted turns in a validated workflow. This is a specific observed path, not a universal cost or performance claim.
More difficult work could escalate to a hosted coding model.
Results could be independently reviewed; failed or unavailable review did not become an approved result.
Git worktrees retained isolated changes while primary working directories remained unchanged.
What I worked through
I’m using this project to test where coding agents are useful and where they need help. The routing, worktree isolation, tests, and separate review all help me inspect what happened when a task goes wrong.
Examples describe engineering methods. Client identities, endpoint identifiers, credentials, and ticket contents are omitted.
I use Qualys to identify systems that need patches, configure deployments, and investigate failed or unclear results. I check installed updates and follow-up scans before considering a finding resolved.
Auditing ESXi and network monitoring, correcting safe configuration issues, validating discovery, and documenting the exceptions that still need investigation.
Production monitoring review
PRTG Network Monitor
VMware ESXi
SSL/TLS
Auto Discovery
Need help with something similar?
Facing a similar technical problem?
This is an example of my experience, not an endorsement from an independent client. If you need similar help, take a look at my services or send me a description of the problem.