Overview
The talk argues that organizational context—not model capability or raw information access—is becoming the central constraint on effective software agents. Human engineers gradually accumulate tribal knowledge through code, architecture documents, incidents, reviews, and conversations, while an agent effectively begins every task as an expert engineer on its first day. Giving the agent a wiki, repository search, or an enormous context window does not solve this problem: it may stop after finding a plausible answer, fail to connect dependencies and past decisions, or become distracted by irrelevant material. Unblocked’s proposed solution is a context engine that synthesizes task-specific understanding from repositories, pull requests, Slack, Notion, architecture documents, and inferred expertise relationships. In demonstrations, this context improved an optimization plan, reduced discovery work, supported evidence-backed answers, recovered prior review guidance, and connected a generated fix to the conversations explaining the underlying regression. The system also exposes its sources so humans can inspect and correct its reasoning. The speaker’s broader thesis is that context quality compounds across planning, implementation, and review: weak assumptions create repeated loops, while relevant context reduces wasted tokens and guides agents toward the right work. As software development moves toward greater automation, agents will require not merely searchable data but continuously synthesized organizational intelligence.
Sections
Core Definitions
Key concepts used to explain why organizational context is necessary for effective agents.
- Context engine: a system that delivers synthesized organizational context to human workers and agents.
- Organizational context: the combined code, intent, conventions, past decisions, discussions, architecture rationale, review history, and expertise relationships needed to understand how work should be performed.
- Satisfaction of search: the tendency to stop investigating after finding one apparently valid result, potentially missing other important evidence.
- Unknown unknowns: relevant constraints or facts that an agent does not know it needs to seek, described in the talk as finding the things that really matter.
- Engineering social graph: a derived map of team and code-review relationships used to identify expertise, clusters, and gaps in codebase coverage.
Strategic Insights
Broader implications derived from the speaker's argument and demonstrations.
- As coding models become more capable, competitive advantage may shift from raw generation quality toward the quality, relevance, and provenance of the organizational context surrounding each task.
- A context engine functions as organizational memory infrastructure: it preserves not only what happened, but why it happened and whose judgment carried weight.
- Context quality is a systems property rather than a prompt-level optimization because errors propagate across planning, execution, debugging, and review loops.
- Traceable sources serve two audiences simultaneously: humans use them to audit claims, while agents use them as navigation points for deeper investigation.
Approaches to Agent Context
Contrasts between raw information delivery and synthesized, task-specific context.
- Repository or wiki access lets an agent search available material; a context engine attempts to identify, connect, and prioritize the material that matters for the task.
- Loading all available material maximizes volume but introduces distraction and token waste; task-specific delivery reduces irrelevant context and supports focused reasoning.
- Without Unblocked, the agent spent more time discovering context and produced a less nuanced optimization plan; with Unblocked, it incorporated historical discussions and proposals while using less time and cost in the demonstration.
- Code reveals current implementation, whereas organizational context supplies intent, conventions, rationale, and historical constraints that are not necessarily encoded in the implementation.
Risks and Failure Modes
Warnings about deploying agents without reliable, focused, and auditable organizational context.
- An agent may accept the first plausible search result and miss other evidence that materially changes the diagnosis or plan.
- Supplying the entire organizational corpus can distract the agent, consume excessive tokens, and weaken task-specific reasoning.
- Incomplete context can produce a wrong plan or false assumptions that compound through repeated implementation and correction loops.
- Generated answers or diagrams may be inaccurate even when presented confidently.
- Expertise signals derived from seniority or review history could over-amplify established voices and underrepresent valid dissenting knowledge.
- Moving toward fully automated agents without organizational context may cause agents to become lost or act against unstated conventions and intent.