The session keeps expanding.
More files, more threads, more agent suggestions. Momentum becomes cognitive load.
Prototype-backed HAI control layer
Sidecar: the agentic control layer behind HAI.
The main agent does the work. Sidecar asks the harder question: is the human still oriented, in scope, and able to verify what just happened?
The control problem
Agent work can look productive while control gets weaker.
Sidecar exists for the moment where AI output is not obviously wrong, but the human has lost the thread: too much context, too many open loops, no clear stop point, and no proof that the work is actually done.
The goal quietly changes.
Each step makes local sense, but the system moves away from the original human intent.
The agent owns too much.
The model starts deciding scope, risk, and completion while the human only reacts.
Green checks replace evidence.
Commits, logs, and tests can look like progress without proving the core requirement.
Sidecar V7
The big attempt: a second agentic layer watching the work.
V7 wrapped Claude Code with hooks, a daemon, safety gates, observation, context injection, slow analysis, and a dashboard. It was not just a tool. It was a serious experiment in making agent collaboration observable.
Hooks
User prompts, tool calls, edits, errors, and phase changes became observable events.
Gates
Risky actions could be blocked or slowed: destructive commands, edit-before-read, weak context.
Signals
Overload, scope creep, planning drift, missing tests, and review gaps became named patterns.
Human
The point was not to automate more. It was to keep the human able to decide, stop, and verify.
The hard lesson
V7 itself became the proof that soft rules are not enough.
Sidecar V7 was powerful, but it also drifted. It grew from a consolidation into a large mechanism-heavy system. That failure is the trust signal: Samuel has already seen how agent-control systems break from the inside.
Agent control needs hard scope anchors, visible state, and independent verification.
A prompt rule is advice. A gate changes behavior. A benchmark is useful only when it checks the thing that actually matters. HAI grew from that distinction.
Sidecar-NG
The answer got smaller and sharper.
Sidecar-NG moved away from a giant daemon toward a local Claude Code control hook: inject one relevant rule at the point of action, and use deterministic gates where risk must not be left to vibes.
01
Read the prompt.
The hook sees what the human is asking before the agent starts moving.
02
Choose one control rule.
Scope, delegation, baseline, verification, or skill routing becomes situational context.
03
Block dangerous tool use.
Destructive commands, force push, secrets, edit-before-read, and weak git flow get hard gates.
What this means for your workflow
You may not need Sidecar. You need the control thinking behind it.
The 4h HAI setup looks at your real agent loop and decides what kind of control belongs there: a watcher rule, a prompt layer, a verifier path, a stop rule, or a simpler human-owned workflow.
Where control breaks.
We locate the point where useful agent work turns into ambiguity or overload.
What stays human.
Judgment, risk, irreversible choices, and acceptance criteria stop being implicit.
The smallest control layer.
Not a dashboard by default. A rule, gate, verifier, or packet may be enough.
What proves it helped.
The workflow gets a concrete follow-up signal instead of another pile of artifacts.
4h HAI setup / Sidecar-fit assessment
Bring one agent workflow where the loop feels out of control.
Leave with a clearer control map: where the agent may act, where the human must decide, what needs verification, and whether a Sidecar-style watcher belongs in the system.