Operationalizable rules for solo parallel coding-agent work. Synthesised from the cognitive-psychology and human-factors literature (Leroy, Mark, Endsley, Parasuraman, Bainbridge) and a 2025-26 practitioner survey (Huntley, Vincent, Cherny, Osmani, Sankalp, Sakasegawa, Daxia, Debois).
The one-week self-audit these rules were supposed to rest on was never logged. Zero daily rows. The numbers below come from the convergence of the literature and the operator survey, not from logged personal behaviour. Treat them as hypothesis-tier, to re-check against a real logging window.
Each operator's stated soft-cap on deeply-attended interactive sessions. The line at x=2 is where independent reports converge before any operator pushes higher.
Everything else is either a fire-and-forget bg job or queued. When a third "I'll just check on this" session appears, triage on the spot: promote one to bg-job (with clear hand-off prompt) or park one explicitly. The binding constraint is the human review pipeline, not the session count.
anchor · Osmani, Vincent, Daxia, Sankalp convergenceFirst message of any session: intent: <verb> <object> — <expected end-state>. Re-orientation budget is 10 seconds. Skipping this is the quiet-session-blackhole default, Bainbridge irony 2 plus Strayer inattentional blindness.
Three lines, typed into the session chat (not a separate notes file): where you left off, what's blocked, next concrete step on return. The cost of skipping isn't measured at the transition, it's residue on the next session for the full intervening window.
anchor · Leroy & Glomb 2018, n=202+66+44Inside a named focus block (deep code work, writing, planning): only "needs input" alerts from the active session stay live. Non-urgent pings defer, everything else queues until the block ends.
anchor · Sakasegawa Patrol Agent patternCompleted tasks produce measurably less residue than interrupted ones, and "perception of incompleteness" is what drives the carryover (Leroy 2009). The 2-minute window is the rough threshold where staying costs less than residue + ready-to-resume note + re-entry. Past two minutes, do rule 3 and switch cleanly. Don't tell yourself "I'm almost done" past that line.
anchor · Leroy 2009 incompleteness driver| Task type | Parallelize? | Cap | Why |
|---|---|---|---|
| Long bg jobs research dispatch, transcribe, summarize URL, dependency upgrade |
yes | none | Fire-and-forget. Review cost is flat at terminal state, no in-flight Level-2 SA needed. |
| Code back-and-forth interactive debug, live refactor, paired implementation |
no | 1 | Continuous Level-2 SA needed. Any switch maxes attention residue, review bottleneck is the binding constraint. |
| Independent research literature scan, product comparison, knowledge-base sweep |
yes-with-cap | 2–3 | Output read at end, low residue if no time pressure. Matches Sankalp pre-implementation pattern. |
| Vault edits filing, knowledge save, link audit, frontmatter sweep |
yes-with-cap | 2 | Mechanical, low review cost. Same-file collisions are likely, isolate by folder (Veske file-boundary rule). |
| Meeting prep 1-1, all-hands, prep doc drafting |
no | 1 | Prevention-framed under time pressure equals max residue (Leroy 2016). Synthesis quality degrades fast under split attention. |
| Multi-repo coordination cross-PR review, merge audit, monorepo restructure |
no | 1 | Cross-file dependencies equal highest review cost. Inattentional-blindness risk highest here, Willison's "parallel agent psychosis" case. |
| Infra / k8s ops prod kubectl, doctl mutations, DB migrations |
no | 1 | High blast radius plus complacency trap (Parasuraman). Demands single-threaded vigilance. This is where automation bias misses subtle wrong choices. |
| Meta-work skill writing, CLAUDE.md edits, vault structure refactor |
yes-with-cap | 2 | Like notes edits with higher recursion risk. Never edit your global agent config from a session that is also shipping live code. |
claude --resume.No defensible number until logged. Provisional working figure: 4 to 6 spawns per focused hour, of which at most 2 are interactive. Re-validate.
End-of-day prompts might help. Without weeks of data on whether they surface patterns, leave it optional and fold into an existing daily or end-of-session routine rather than building a separate ritual.
Switch costs scale with task dissimilarity (2024 parametric extension). Two backend refactor sessions on adjacent files might be cheaper to attend in parallel than one backend plus one creative-writing. The cap could be context-sensitive, but the practitioner data doesn't separate these. Open.