Parallelism is seductive
The first thrill of agents is parallelism. Ten workers can research, draft, refactor, summarize, compare, and test at once.
The trap is integration debt.
Parallel execution does not automatically create parallel progress. It creates more artifacts, more branches, more claims, more decisions, and more surfaces to reconcile. Someone has to maintain the plot.
That someone is the individual operator.
Portfolio management, not agent chaos
Treat agents as a portfolio of bounded workstreams. Each run should have a scope, source base, risk level, checkpoint, output path, review burden, and decision attached.
The operator is not there to micromanage every token. They are there to manage dependencies, sequence, risk, and synthesis.
A useful portfolio board tracks:
- Task: the bounded unit of work;
- Worker/tool: which agent, script, or person is responsible;
- Source base: what the worker is allowed to rely on;
- Risk level: low, medium, high, or external-facing;
- Checkpoint: when the operator will inspect progress;
- Output path: where the artifact lands;
- Reviewer: who decides whether it is acceptable;
- Decision needed: accept, revise, merge, escalate, discard;
- Blocker: what prevents completion;
- Memory update: what should be preserved after completion.
If the board is invisible, the work system lives in chat scrollback. Chat scrollback is not an operating system.
The synthesis problem
The most common portfolio failure is many locally successful runs that produce globally incoherent work.
One agent writes the plan. Another rewrites the examples. A third updates the tracker. A fourth mirrors files. Each run reports success. Then the operator discovers the title changed in one place, the status changed in another, the source pack drifted, and the next step no longer matches the actual state.
No individual worker failed. The portfolio failed.
Synthesis is a first-class task, not an afterthought.
Trust ladders
Not every agent run deserves the same autonomy. Use a trust ladder:
- Draft only: the worker can propose but not change files.
- Edit inside a bounded folder: reversible local changes only.
- Run tests or checks: execution allowed, but no external side effects.
- Update trackers/mirrors: allowed after verification.
- External or irreversible actions: human review required.
Autonomy should expand because reliability was demonstrated, not because the operator got tired.
Portfolio review checklist
At the end of a multi-agent cycle, ask:
- Are all outputs pointing to the same decision?
- Did any worker silently change scope?
- Are statuses, trackers, and mirrors consistent?
- Which artifacts should be merged, discarded, or sent back?
- What did the portfolio teach the system?
- Did parallelism actually save time after review and synthesis?
More workers can make you slower if you do not design the queue. The agent portfolio is leverage only when the operator can still see the whole.
