Agent vs Human Tradeoffs

Decision matrix когда agent cost-effective, когда human essential. Complements Economics-Framework (quantitative) с qualitative judgment.

Core principle

Not “automate everything” и not “keep all manual”. Deliberate choice per function based on properties task / decision.

Manifesto vision: 30-50 agents, Founder strategist-only. But “strategist- only” ≠ 100% automation. Founder keeps:

• Strategic bets
• Relationships
• Crisis leadership
• Codex judgment
• Irreversible decisions

Everything else — candidate for agent.

When agents cost-effective

1. High-frequency structured tasks

• Research tasks, data gathering
• Content drafting (not final publishing)
• Standard analysis (comparisons, summaries)
• Routine communications (internal)
• Scheduled checks / monitoring

Why: repetition amortizes agent development cost. Structure allows playbook.

2. Pattern-recognition work

• Competitor monitoring (detecting changes)
• Market signal aggregation
• Anomaly detection
• Categorization / tagging

Why: LLMs excel at pattern matching. Faster than humans при equal accuracy.

3. Cognitive grunt work

• Summarization long content
• Data entry / transformation
• Format conversion
• Quality check (per Agent-Judge pattern)

Why: humans bored by это → worse quality than machines. Machine neutral.

4. Parallel-able tasks

• Research 10 competitors simultaneously
• Check 20 sources concurrently
• Generate multiple variants

Why: parallelism cheap for agents, expensive for humans (need coordination).

5. 24/7 availability

• Monitoring (system, external events)
• Async task queues
• Customer-facing FAQ responses

Why: humans sleep. Agents don’t. For tasks requiring continuous availability, agent cheaper.

6. Reversible work

• Drafts (not published)
• Memory writes (can be corrected)
• Task queue management

Why: mistakes recoverable. Agent autonomy safe.

When humans essential

1. Novel / unprecedented situations

Agent training / memory doesn’t cover new situations well. Examples:

• New market entry decisions
• First-of-kind partnership
• Unexpected market events
• Crisis response

Why: LLMs confident в novel situations but often wrong (hallucinate patterns). Human judgment rougher but better-calibrated на absence of data.

2. Relationship-building

• Founder-to-founder conversations
• Negotiation (non-scripted)
• Community building
• Reputation management

Why: trust built between humans. Agent “simulated” relationship brittle.

3. Strategic bets

• Pivot decisions
• Major resource allocations
• Multi-year direction setting

Why: long-horizon optimization requires context agents don’t have + responsibility humans accept differently.

4. Codex judgment

• Interpretation ambiguous rules
• Exception requests
• Ethical edge cases

Why: Codex values require human values-application. Agents compliance-check but don’t arbitrate values.

5. Irreversible decisions

• Public communications (at scale)
• Contractual commitments
• Burn bridges decisions

Why: no undo. Cost of mistake too high.

6. Human-required by nature

• Interviews / interrogation (investigations)
• Witnessing / verification (legal, compliance)
• Signatures / attestations
• Crisis press conferences

Why: institutional / legal requirements.

7. Trust-building с stakeholders

• Investor conversations (especially negative news)
• Board reporting
• Regulator engagement

Why: stakeholders want humans accountable.

Hybrid patterns (both)

Many tasks benefit from human-agent hybrid, not either/or:

Agent-proposes, human-decides

Agent generates options, analysis, recommendation. Human chooses.

Examples:

• Strategy options → Founder picks
• Hiring recommendations → HRBP decides
• Pricing analysis → Leadership sets

Agent-drafts, human-edits

Agent writes v1, human refines.

Examples:

• Content (blog, newsletter) — agent draft, human brand voice final
• Email communication — agent structure, human tone
• Documents — agent Research + template, human final polish

Agent-executes, human-reviews

Agent does work, human samples для quality.

Examples:

• Research outputs — Agent-Judge + periodic human sample
• Data entry — agent does, spot-check human
• Automated decisions — rolling sample review

Human-escalates-to-agent

Less common. Human feels task mundane → punts к agent.

Examples:

• Scheduling / calendar management
• Email triage
• Routine follow-ups

Functional ranking for Synth Nova

[applicable] — текущая оценка automation feasibility Synth Nova functions:

High automation potential (>70% of work to agent)

• Market / competitor research → Agent-MarketResearcher
• Niche evaluation (data gathering) → Agent-NicheEvaluationDirector
• Task coordination → Agent-CEO
• Quality checks → Agent-Judge
• Content drafting (future)
• Data monitoring (future)

Medium automation potential (40-70%)

• Business analysis — agent analysis, Founder interpretation
• Decision support — agent options, human chooses
• Communication drafting — agent drafts, Founder tone

Low automation potential (<40%)

• Strategic bets — Founder judgment + Chamber advisory
• External relationships — Founder-driven
• Legal / regulatory — if arises, human + outside counsel [future]
• Crisis response — human-led

Not automated (Founder only)

• Fundraise decisions [future]
• Board interactions [future]
• Existential decisions (pivot, shutdown)
• Codex amendments
• Individual Agent-CEO direction setting

Tradeoff table

Comparison по dimensions:

Dimension	Agent	Human
Speed (per task)	Seconds-minutes	Minutes-hours
Throughput (parallel)	High	Low
Cost per task	$0.001-$5	$50-$500+
Consistency	High (same playbook)	Variable
Novel situations	Poor (unless in training)	Better (real-time reasoning)
Relationship building	None	Essential
Judgment на values	Compliance check only	Values arbitration
24/7	Yes	No
Learning curve	Days (prompt tuning)	Months
Accountability	None (tool)	Full
Trust от stakeholders	Low	High
Creativity на novel problems	Mixed	Generally better

Dynamics over time

Frontier shifts. What’s “human only” today may become “agent + human” tomorrow.

• LLMs improving (frontier moves outward)
• Tooling improves (agent capabilities expand)
• Learned patterns expand automation scope

Rule: revisit tradeoffs quarterly. Don’t lock-in decisions.

Risk factors

Over-automation

Automating too much → brittle system, lose flexibility, dependency on models.

Signs: человек теряет intuition на domain. Failure modes not noticed. Cost cutting trumps quality.

Under-automation

Keeping too much manual → Founder bottleneck, doesn’t scale, burnout.

Signs: Founder overloaded. Tasks pile up. Manifesto vision (30-50 agents) not progressing.

Wrong automation

Automating wrong thing → ROI negative, expensive lesson.

Signs: Agent cost ≈ human cost. Success rate <85%. High HITL rate. Frequent rollbacks.

Synth Nova application

Current phase (MVP, ~5-10 agents):

• Prioritize: Q1 and Q2 categories (high / medium feasibility, reversible)
• Defer: strategic / relationship tasks
• Keep manual: Founder direct work, external relationships

Next phase (post-MVP):

• Expand agent tiers — domain specialists
• Add policy-sensitive agents (legal / compliance [future])
• Introduce Chamber for strategic decisions
• Start outcome labeling discipline (Process-OutcomeLabeling)

Long-term (Manifesto 12-month vision):

• 30-50 agents across domains
• Founder strategic-only
• Board / formal governance structure [future]
• International expansion если applicable

Связанные документы

• Economics-Framework — quantitative sizing
• Manifesto — vision 30-50 agents, principles
• Rules-AgentDecisionBoundaries — what agent may decide
• Rules-Criticality — L1-L5 classification
• Reference-Org-Blueprint — section §10 framework
• Synth-Nova-Overview
• MVP-Phase
• Agent-CEO, Agent-Judge, _Roles-Index
• Process-HypothesisValidation
• Build-Measure-Learn