Agentic AI in the Enterprise (Praveen Akkiraju, CXOTalk)
enterprise-aiagentic-aiharnessjagged-intelligencetoken-maxingbuild-vs-buycio
Agentic AI in the Enterprise (Praveen Akkiraju, CXOTalk)
Praveen Akkiraju (Managing Director, Insight Partners) talks with Michael Krigsman about the state of enterprise agentic AI in 2026. The cleanest articulation of "the agent IS the harness" plus a usable bounded/unbounded framework for autonomy decisions.
Key claims
- Model evolution arc: late 2024 GPT "4 model" introduced reasoning → DeepSeek changed economics → 2025 = year of agents (computer use, browser use) → 2026 inflection point. Sophisticated enterprises now run 1,000+ agents in production.
- "The agent IS the Harness (LLM Agents)." LLMs are stateless; the harness (tools, context, memory, guardrails) is what makes work possible. "The actual constraints and steering of the agent has to happen in the harness."
- Jagged Intelligence is real: same model can produce SOTA code AND make trivial mistakes. Implies you must do harness homework per use case.
- Where agents deliver real value today: customer support (
$500M revenue), coding ($3B revenue), enterprise search. Verticals: legal, healthcare, tech-forward enterprises. - Token Maxing: enterprises blow through annual AI budgets in 90 days. "As many tokens as you provide will get consumed as quickly as possible." Goldman Sachs study cited (uncorroborated).
- Negligence/observability: agents can pick wrong tools, take inputs astray, fall to prompt injection. Design pattern: ephemeral sandboxes (e.g. E2B) that execute agent-written code against guardrails before production push. Observability must be at every step, not just the final output. Errors compound in multi-agent architectures.
- Data access boundaries: pick the right interface — APIs give granular OAuth control, MCP is more abstracted. Use
.md"policy files" to instruct agents about PII / sensitive data handling. - Agent vs agent-native software (the "SaaS apocalypse" question): hybrid future. Bolted-on agents on legacy software lose; thoughtful insertion of agentic steps into existing workflows wins (Stampli AP-reconciliation example). Bottoms-up agent-native building only justified where there's strong data platform leverage (security ops, IT ops, customer service ops).
- Bounded vs Unbounded Tasks framework for autonomy:
- Bounded (math, coding, deterministic workflows) → autonomy possible, sometimes already at "AGI" levels per labs
- Unbounded (supply chain across SKUs/geos/suppliers) → harness design dominates; full autonomy distant
- Three things shape the Human in the Loop dial: (1) workflow nature (bounded↔unbounded), (2) regulatory/compliance environment, (3) harness quality.
- Fix data first, or start now? Start now. You can't afford to wait. Computer-use and browser-use let agents work over messy data like a human would. Different from ERP/CRM era — this is moving too fast for big-bang transformations.
- Hiring shifts: roles are mashing together. Software engineer + PM + forward-deployed engineer in one person. Iteration loops are faster, silos are breaking down.
- OpenClaw moment ("earlier this year"): first open-source personal-agent framework that actually worked — right context, memory, tooling. "Truly blew people's minds."
- Build vs Buy = front-end vs back-end (Build vs Buy (Agents)):
- Front-end / standardized workflows → buy (customer support, finance reporting)
- Back-end / data-heavy / industry-specific → build (using frameworks)
- Token cost reality: software is now variable-cost. Token + tool + platform cost must be weighed against ROI. Models are getting more efficient (Opus 4.5 → 4.7; GPT 5.5).
- Pricing innovation: fractional FTE pricing emerging — agents priced as partial headcount, not seats.
- General-purpose agent libraries are not yet trustworthy. Be careful deploying third-party agents. Mythos exception (security): one Mozilla scan found 271 Firefox bugs, ~3 high severity.
Cross-source resonance
- Jagged Intelligence also Karpathy's framing in Andrej Karpathy on Agentic Engineering (Sequoia AI Ascent).
- Harness = agent matches Boris Cherny's view in Boris Cherny on Coding Is Solved (Sequoia AI Ascent) — though Boris predicts harness will matter less as models improve, while Praveen says today it's the determinant. Tension worth tracking on Harness (LLM Agents).
- OpenClaw reference reinforces it as the canonical personal agent.
- "Ephemeral sandboxes" = same idea as Boris Cherny's batch/loop ephemeral execution and the Printing Press real-Chrome-session-per-CLI approach.